Genetic population structure constrains local adaptation and probability of parallel evolution in sticklebacks

Abstract

Repeated and independent adaptation to specific environmental conditions from a shared ancestral pool of standing genetic variation (parallel evolution) is common. However, if standing genetic variation is limited, local adaptation may be restricted to different and potentially less optimal solutions (convergent evolution), or prevented from happening altogether. Using a quantitative trait locus (QTL) mapping approach, we identified genomic regions responsible for the repeated pelvic reduction (PR) in three crosses between nine-spined stickleback (NSS) populations expressing full and reduced pelvic structures (PS). In one cross, PR mapped to linkage group 7 (LG7) containing the gene Pitx1 known to cause PR. In the other two crosses, PR was polygenic and attributed to ten novel QTL, of which 90% were unique to specific crosses. Next we screened the genomes of individuals from 27 different NSS populations for deletions in the Pitx1 regulatory element (Pel), which is known to be associated with repeated PR in three-spined sticklebacks (TSS). This revealed deletions only in the population where PR mapped to LG7. Using individual-based forward simulations parametrised with empirical data on NSS and TSS population structures, we show that access to the same ancestral variation promotes parallel evolution (as seen in TSS), whereas lack thereof restricts local adaptation and instead promotes emergence of convergent genetic architectures (as seen in NSS). Hence, the results predict that with an increasing degree of population subdivision, parallel phenotypic evolution becomes increasingly non-parallel at the genetic level and restricts local adaptation to less optimal solutions.

Introduction

Failure to evolve in response to changing environmental conditions may lead to extirpation or even extinction (Orr and Unckless 2008). If the heritable variation necessary for adapting to environmental change already exists in the form of standing genetic variation, genetic adaptation may proceed swiftly, at least compared to the time it would take for populations to adapt from novel mutations (Orr and Unckless 2008; Barret and Schluter 2008; Thompson et al. 2019). Therefore, when exposed to novel yet similar environments, populations derived from the same ancestral population – hence carrying the same pool of alleles – can often be expected to respond to similar selection pressures in a similar fashion. The resulting evolution of parallel genetic adaptations is known as parallel evolution (Arendt and Reznick 2007; Schluter and Conte 2009; Elmer and Meyer 2011; Stern 2013; Conte et al. 2015; Bolnick et al. 2018; Barghi et al. 2019; Hermisson and Pennings 2017). However, in genetically highly structured species, potentially advantageous rare alleles may be lost due to founder events and random genetic drift, preventing adaptation. Alternatively, adaptation to given selection pressures might be more easily gained by allelic substitutions in alternate loci (convergent evolution) influencing the same polygenic trait, even if they may significantly differ in their fitness effects (Cohan 1984; Merilä 2013, 2014; Rosenblum et al. 2014). Thus, the demographic history of populations may play a central role in determining the likelihood of parallel evolution.

The three-spined stickleback (Gasterosteus aculeatus) is a widely used model system to study adaptive evolution in the wild (Bell and Foster 1994; Gibson 2005). Its ability to rapidly adapt to local environmental conditions has often been shown to stem from a global pool of ancestral standing genetic variation (Schluter and Conte 2009; Jones et al. 2012; Terekhanova et al. 2014, 2019). The nine-spined stickleback (Pungitius pungitius) has been emerging as another model system for the study of repeated evolution in the wild (Merilä 2013). In general, it differs from the three-spined stickleback by having smaller effective population sizes (N_e), reduced gene flow in the sea, and a tendency to occur in small landlocked ponds in complete isolation from other populations (Shikano et al. 2010; DeFaveri et al. 2012; Merilä 2013). Given their contrasting population demographic characteristics, three- and nine-spined sticklebacks can thus be expected to occupy different parts of a parallelism-convergence continuum with respect to local adaptation.

Regressive evolution of the pelvic complex has occurred in three (viz., Gasterosteus, Pungitius, and Culaea) of the five recognised stickleback genera since the last glacial period (reviewed in Klepaker et al. 2013). While marine populations of three- and nine-spined sticklebacks usually have complete pelvic structures with fully developed pelvic girdles and lateral pelvic spines, partial or even complete pelvic reduction is common in freshwater populations (Blouw and Boyd 1992; Shapiro et al. 2004, 2006; Herczeg et al. 2010; Klepaker et al. 2013). Several factors may contribute to this, including the absence of gape-limited predatory fish and limited calcium availability, as well as the presence of certain insect predators (Reist et al. 1980; Reimchen et al. 1983; Giles 1983; Bell et al. 1993; Karhunen et al. 2013; Chan et al. 2010). Collectively, sticklebacks provide an important model system to study genetic mechanisms underlying the adaptive parallel pelvic reduction at both intra- and inter-specific levels, under a wide range of population demographic settings. However, studies of parallel patterns of marine-freshwater divergence in nine-spined sticklebacks are still scarce (Herczeg et al. 2010; Wang et al. 2020), especially at the genetic level, precluding any comprehensive and conclusive comparison of the two species.

The genetic basis of pelvic reduction in the three-spined stickleback is well understood. Quantitative trait locus (QTL) mapping studies have identified a single chromosomal region containing the gene Pituitary homeobox transcription factor 1 (Pitx1) that explains more than two thirds of the variance in pelvic size in crosses between individuals with complete pelvic girdles and spines, and pelvic-reduced individuals (Cresko et al. 2004; Shapiro et al. 2004; Coyle et al. 2007). Pelvic loss in the marine-freshwater three-spined stickleback model system is predominantly caused by expression changes of the Pitx1 gene due to recurrent deletions of the pelvic enhancer (Pel) upstream of Pitx1 (Chan et al. 2010; Xie et al. 2019). However, in benthic-limnetic and lake-stream pairs of three-spined sticklebacks, the genetic architecture of pelvic reduction appears to be more diversified (Peichel et al. 2001, 2017; Deagle et al. 2012; Stuart et al. 2017). In the three independent nine-spined stickleback QTL studies available to date, Pitx1 in linkage group (LG) 7 was also identified as a major cause for the reduction in pelvic structures in one Canadian (Shapiro et al. 2006) and one Finnish (Shikano et al. 2013) population. Another large effect region in LG4 was found to be associated with pelvic reduction in an Alaskan population (Shapiro et al. 2009). Similar to the three-spined stickleback, pelvic spine and pelvic girdle sizes are strongly correlated in the population from the Finnish pond Rytilampi, since Pitx1 controls both phenotypes (Shikano et al. 2013; Fig. 1 and Supplementary Table 1). In contrast, a considerable amount of phenotypic variation with respect to these traits and their inter-correlations exists among different freshwater pond populations in northern Europe (Herczeg et al. 2010; Karhunen et al. 2013, 2014; Fig. 1). Given their high heritability (Blouw and Boyd 1992; Leinonen et al. 2011), the lack of correlation between spine and girdle lengths suggests that they can be independently controlled by different QTL. Thus, the genetic underpinnings of pelvic reduction in nine-spined sticklebacks could be more diversified than those in the marine-freshwater three-spined stickleback model system (Merilä 2013, 2014).

Figure 1. Summary of previously published nine-spined stickleback pelvic phenotypes from the wild and common garden experiments.

Violin plots depict relative (to standard body length) pelvic spine and girdle lengths according to population and habitat type. Numbers in the top panel show P-values for Pearson’s product moment correlation between relative pelvic spine and relative pelvic girdle lengths, and the bottom panel the respective squared correlation coefficients for each population/cross. “na” indicates data with no variation in spine or girdle lengths. Further details can be found in Supplementary Fig. 1 and Supplementary Table 1. Phenotypic data for F₂ individuals from the current study are also presented (HEL x RYT, HEL x BYN and HEL x PYÖ), are also shown for comparison. Details of sample locations and data collection can be found from Herczeg et al. (2010) and Karhunen et al. (2013). Data from common garden experiments is indicated by the suffix “_CG” in x-axis labels.

To investigate the possible genetic heterogeneity underlying pelvic reduction in different nine-spined stickleback populations, we analysed data from three F₂ generation inter-population crosses (279–308 progeny per cross) between pond and marine individuals to identify QTL responsible for pelvic reduction in different ponds. One of these populations was the previously studied Rytilampi cross (earlier analysed only with microsatellites, Shikano et al. 2013), now re-analysed along with two new populations (Bynästjärnen and Pyöreälampi) using > 75 000 SNPs. This data was subjected to a cutting-edge mapping approach (Li Z. et al. 2017, 2018) that can provide more information on the source of the QTL effects than has been previously possible. In addition, to assess the extent to which Pel is responsible for pelvic reduction in the nine-spined stickleback, we screened the whole genomes of individuals from 27 wild populations for deletions in the genomic region spanning the Pel element and the Pitx1 gene. Finally, we used empirical data to estimate realistic levels of population structuring and genetic isolation by distance (IBD) in marine populations of three- and nine-spined sticklebacks, and then used individual-based forward simulations to test how these factors affect the prevalence of genetic parallelism when local adaptation proceeds from standing genetic variation.

Materials and Methods

Fish collection, crossing, and rearing

For the QTL crosses, three different marine F₀ generation females from the Baltic Sea (Helsinki, Finland; 60°13′N, 25°11′E) were crossed with a freshwater F₀ generation male from either Bynästjärnen (64°27′N, 19°26′E), Pyöreälampi (66°15′N, 29°26′E) or Rytilampi (66°23′N, 29°19′E) ponds. Fish crossing, rearing, and sampling followed the experimental protocol used in the earlier study of the Rytilampi population (Shikano et al. 2013; Laine et al. 2013). For Rytilampi, the F₀ generation fish were artificially mated in the lab during the early breeding season of 2006 (Shikano et al. 2013), and the resulting full-sib F₁-offspring were group-reared in aquaria until mature, as explained in Shikano et al. (2013). Two randomly chosen F₁ individuals were mated repeatedly (seven different clutches) to produce the F₂ generation between September and October 2008. The offspring were reared in separate aquaria. The same procedure was followed for Pyöreälampi (F₀ mating: Jun 2011; F₁ mating: Jul–Sep 2012; F₂ rearing: Jul 2012–Apr 2013; 8 different clutches) and Bynästjärnen (F₀ mating: Jun 2011; F₁ mating: Nov 2013– Jan 2014, F₂ rearing: Nov 2013–Aug 2014; 6 different clutches). The F₂ fish were euthanized at 187, 238 and 238 days post-hatch for Rytilampi, Pyöreälampi and Bynästjärnen, respectively. At this point, the fish were on average 52.3 mm in standard length (Rytilampi = 48.6 mm; Pyöreälampi = 52.3 mm; Bynästjärnen = 53.6 mm), and weighed on average 1.34 g (wet weight; Rytilampi = 1.07 g; Pyöreälampi = 1.49 g; Bynästjärnen = 1.44 g). In total, 308, 283 and 279 F₂ offspring were available for analyses from Helsinki × Bynästjärnen (HEL × BYN), Helsinki × Pyöreälampi (HEL × PYÖ) and Helsinki × Rytilampi (HEL × RYT) crosses, respectively.

The experiments were conducted under licenses from the Finnish National Animal Experiment Board (#STH379A and #STH223A). Experimental protocols were approved by the ethics committee of the University of Helsinki, and all experiments were performed in accordance with relevant guidelines and regulations.

Morphological data

To visualise bony elements, all of the F₂-progeny were stained with Alizarin Red S following Pritchard and Schluter (2001). Pelvic spine and girdle lengths from both sides of the body, as well as standard body length, were measured with digital calipers to the nearest 0.01 mm. Although it is known that the left-right asymmetry of the pelvic girdle is heritable in sticklebacks (Blouw and Boyd 1992; Bell et al. 2007; Coyle et al. 2007), we did not specifically study this. To reduce the number of tests, the mean of the left and the right side measurements was used (analyses conducted for the two sides separately always yielded qualitatively similar results as the tests conducted with the averages; results not shown). All measurements were made by the same person twice; the repeatability (R; Becker 1984) ranged between 0.80 and 0.84 for girdle lengths, and between 0.98 and 0.99 for spine lengths. The QTL analyses were performed on both absolute and relative (scaled to the total body length) trait values, but for all of the analyses that compared phenotypic data between populations (which also differ in body sizes), only relative trait values were used. Previously published phenotypic data from 19 wild populations were obtained from Herczeg et al. (2010) and Karhunen et al. (2013). These included data on pelvic spine and girdle lengths of wild-collected individuals from ten pond populations (Abbortjärn = ABB, Bolotjone = BOL, Karilampi = KAR, Kirkasvetinen lampi = KRK, Mashinnoje = MAS, Lil-Navartjärn = NAV, Hansmytjärn = HAN, Rytilampi = RYT, Bynästjärnen = BYN), four lake populations (Iso Porontima = POR, Riikojärvi = RII, Joortilojärvi = JOR, Västre-Skavträsket = SKA) and five marine populations (Fiskebäckskil = FIS, Trelleborg = TRE, Bölesviken = BÖL, Helsinki = HEL LEV = Levin Navolak), as well data on common garden-reared F₁ generation individuals from two marine (HEL & LEV) and two pond (BYN & PYÖ) populations. Visibly broken spines were treated as missing data. A map showing the geographic location of these populations is provided in Supplementary Fig. 1.

Genotyping and linkage map construction

The RAD sequencing protocol used to obtain the SNP data was the same as in Yang et al. (2016) and Li Z. et al. (2017). In short, genomic DNA was extracted from ethanol preserved fin clips using the phenol-chloroform method (Taggart et al. 1992). DNA was fragmented with PstI restriction enzyme and the resulting 300 to 500 bp long fragments were gel purified. Illumina sequencing adaptors and library specific barcodes were ligated to the digested DNA fragments, and the barcoded RAD samples were pooled and sequenced on 24 lanes of the Illumina HiSeq2000 platform with 45 bp single-end strategy. RAD library construction and sequencing were conducted by BGI HONGKONG CO., LIMITED. After eliminating adapters and barcodes from reads, a quality check was done using FastQC (http://www.bioinformatics.bbsrc.ac.uk/projects/fastqc/).

Linkage mapping for the three crosses was conducted using Lep-MAP3 (LM3; Rastas 2017), as described in detail in Li H. et al. (2018). LM3 can infer the parental/grandparental phase based on dense SNP data, which allowed us to utilise the four-way cross QTL mapping method described below. Input data were generated by first mapping individual fastq files to the nine-spined stickleback reference genome using BWA mem (Li H. 2013) and SAMtools (mpileup; Li H. et al. 2009), followed by pileupParser.awk and pileup2posterior.awk scripts from the LM3 pipeline using default parameters.

The mapping was carried out following the basic LM3 pipeline and by combining the three crosses: parental genotypes were called by taking into account genotype information on offspring, parents and grandparents (module ParentCall2). Markers segregating more distortedly than what would be expected by chance (1:1,000 odds) were filtered out (module Filtering2). Loci were then partitioned into chromosomes (modules SeparateChromosomes2 and JoinSingles2) yielding 21 linkage groups with > 75 000 markers assigned to these groups. Finally, the markers were ordered within each linkage group with the module OrderMarkers2, removing markers that were only informative in either the mother or father, respectively. This created two maps for each chromosome, one having more maternal markers and the other having more paternal markers, with an average of ⅔ of the markers shared between the two. Constructing two maps this way removes the effect of markers that are informative only in one parent, as markers informative in different parents are not informative when compared against each other. The phases were converted into grandparental phase by first evaluating the final marker orders and then matching the (parental) phased data with the grandparental data, inverting the parental phases when necessary. Orphan markers from map-ends based on scatter plots of physical and map positions were manually removed.

Dimensionality reduction by linkage disequilibrium network clustering

In QTL mapping, it is essential to correct for multiple testing in order to reduce the rate of false positives. Moreover, in large genomic datasets, physically adjacent SNPs – particularly those from experimental crosses – are often in linkage disequilibrium (LD), i.e. correlated. Since a group of SNPs in high LD explains similar amounts of genetic variation in a given trait, it is reasonable to apply a dimensional reduction procedure before QTL-mapping to exclude the redundant information from the data. Here we used linkage disequilibrium (LD) network clustering (“LDn-clustering”) and PC regression as dimensionality reduction tools prior to single-locus QTL mapping (Li Z. et al. 2018). The first step of this approach involves an extension of a method developed by Kemppainen et al. (2015), which uses LD network analysis for grouping loci connected by high LD. The second step involves principal component analysis (PCA) as a method for dimensionality reduction in each cluster of loci connected by high LD (“LD-clusters”). This was achieved by the function “LDnClustering” in R-package “LDna” (v.0.64; Li Z. et al. 2018). The method used here differs slightly from the original method described in Li Z. et al. (2018), and was implemented as follows. Initially all edges (representing LD-values as estimated by r² using the function “snpgdsLDMat” from the R-package “SNPRelate”, Zheng et al. 2012) below the LD threshold value of 0.7 were removed. This resulted in many sub-clusters in which all loci were connected by at least a single edge. Second, for each sub-cluster, a complete linkage clustering was performed, where a cluster is defined by its smallest link. Starting from the root, additional sub-clusters were found recursively, where the median LD between all loci was > 0.9. This recursive step was not used in the original implementation, but was later found to increase computational speed and reduce the number of single locus clusters, i.e. result in more efficient complexity reduction. To facilitate computational speed, we only considered SNPs no further than 2000 SNPs away from each other (rather than considering all pairwise values within a chromosome at a time; parameter w2 = 2000), but nevertheless analysed all SNPs from each cluster at a time (rather than performing LD clustering in windows). Thus, each resulting LD-cluster represents sets of physically adjacent and highly correlated loci. We performed PCA regression based on each LD-cluster in which individuals are separated according to their LD-cluster multi-locus genotypes. Since the loci are highly correlated, most of the genetic variation (>90%) between individuals is explained by the first principal component. For each LD-cluster, we then replaced all the SNP genotypes by the position of the individuals along the first PC axis after removing the sex-linked loci (Pearson’s correlation coefficient >0.95 between the PC coordinates and sex). The input data for the LDn-clustering was comprised of the original co-dominant SNP data with individuals from all three crosses pooled. A custom R-code used for this dimensionality reduction is available from DRYAD (xxx).

QTL mapping: four-way crosses model

In some circumstances, such as a four-way cross (Xu 1996), F₁ hybrids of two heterozygous parents (Van Ooijen 2009), and an outbred F₂ design (Xu 2013a), it is possible that up to four different alleles – A and B from the dam, and D and C from the sire – segregate in the population. In such cases, a linear regression model for QTL analysis of the outbred F₂ data (Li Z. et al. 2018) is defined by: where y_i is the phenotype value of individual i (i=1,…,n), x_dij, x_sij, z_ij are genotypes coded as

(Xu 2013b), β₀ is the parameter of the population mean, β_dj is the substitution effect of alleles A and B of the dam (i.e. the grandfather in F₀) at the locus j (j=1,…,p; p is the total number of SNPs), β_dj is the substitution effect of alleles C and D of the sire (i.e. the grandmother in F₀), γ_j is the dominance effect, and ε_i is the residual error term mutually following an independent normal distribution.

The model (1) requires the knowledge of the grandparental phase (produced by LM3) with the benefit that the source (viz. F₀ female or F₀ male) of the observed QTL effect can be inferred, as described in more detail in Supplementary File 1. A multiple correction on the basis of permutation tests was further conducted to control for false positives due to multiplicity (Li Z. et al. 2017) with 10 000 replicates.

Estimating the proportion of phenotypic variance explained by SNPs

The overall proportion of phenotypic variance (PVE) explained jointly by all SNPs (an approximation of the narrow sense heritability) was obtained using LASSO regression (Tibshirani 1996), which incorporates all the SNPs into a multi-locus model: where the l₁ penalty term shrinks the regression parameters towards zero; all other symbols are defined in the same way as in Equation (1).

Following Sillanpää (2011), the PVE can be estimated by the formula: where β _j; is the effects of the SNPs, and is the LASSO residual variance estimated by a cross-validation-based estimator introduced by Reid et al. (2016). The PVE explained by each linkage group was estimated on the basis of LASSO estimates using the following formula: where represents all the effects estimated by the LASSO of the SNPs which do not belong to a set of SNPs (e.g. to a chromosome/linkage group). A similar approach was used to estimate the contribution of grandparental alleles and to evaluate the dominance component by treating the coding systems [x_dij, x_sij, z_ij] (2) as different groups of SNPs. A custom R-code for PVE estimation is available from DRYAD (XXX).

Scanning for Pel deletions in full-genome sequence data

A minimum of 20 samples from populations RYT, MAS, BOL, BYN and PYÖ, and 10– 31 samples from an additional 22 populations (from Northern Europe and USA; Supplementary Fig. 1 and Supplementary Table 2) were sequenced to 10× coverage by BGI HONGKONG CO., LIMITED. Reads were mapped to the nine-spined stickleback reference genome (Varadharajan et al. 2020) with BWA mem (Li H. 2013), and site-wise sequencing coverages were computed with SAMtools (depth; Li H. et al. 2009). Relative sequence depths across the H2afy-Pitx1 intergenic region were estimated for the five focal populations by first computing the median depths for 1000 bp sliding windows, and then normalising these by the median depth for the full intergenic region. The Pel-2.5kb^SALR region was extracted from the original BAC contig (GenBank accession number ‘GU130433.1’) and mapped to the nine-spined stickleback intergenic region with minimap2 (Li H. 2018). The sequencing depths for the Pel-2.5kb^SALR were normalised by dividing the mean depths of the Pel region by the mean depths of the full intergenic region. Gene annotations and relative sequencing depths (average and confidence intervals) were computed and visualised using R-package “Gviz” (Hahne and Ivanek 2016). Lastly, we scanned the literature for genes that are known to affect pelvic and hind limb development and searched for potential matches in the nine-spined stickleback genome (Varadharajan et al. 2020) in regions where significant QTL were found. The custom R-code for these analyses is provided in DRYAD (XXX).

Proof of concept simulations

To obtain empirical data to inform our simulations, we analysed a subset of 4326 SNPs from 236 nine-spined and 98 three-spined sticklebacks from 54 Fennoscandian populations. This data was extracted from previously available RAD-seq and full genome sequence data as described in Supplementary File 2. The purpose of this was to provide a ballpark estimate of IBD and degree of freshwater-freshwater genetic differentiation for the simulations. For IBD analyses, the pairwise genetic distance between populations were regressed against the geographic distance separating the populations. We estimated geographic distance in the World Geodetic System 84 reference ellipsoid (i.e. point distance when taking into account the curvature of the planet) using the function “pointDistance” from the R-package “raster” (Hijmans and van Etten 2012). Genetic distances were calculated as F_ST/(1-F_ST) (linearised F_ST), with F_ST estimated as shown above for all pairwise population comparisons. Slope and confidence intervals of the regression lines were estimated by bootstrapping SNPs (1000 bootstrap replicates). As the geographic distances between the marine populations in the simulations were on an arbitrary scale (and uniform between all adjacent marine populations), we scaled the geographic distance in the simulated data such that the slope for a regression line (mean from bootstrapping) between linearised F_ST and geographic distance for the empirical and simulated data would be the same.

The repeated local adaptation in independently colonised freshwater stickleback populations is widely considered to be due to selection on standing genetic variation available in the sea, which in turn is maintained by recurrent gene flow from previously colonised freshwater populations (the “transporter hypothesis”; Schluter and Conte 2009). We used individual-based forward simulations (QuantiNemo2; Neuenschwander et al. 2018) to investigate how our estimated differences in population structure between nine- and three-spined sticklebacks under the transporter hypothesis influence the prevalence of parallel evolution, following post-glacial colonisation of freshwater habitats from the sea. For both species, the sea populations were simulated by a stepping-stone model comprised of ten sub-populations with carrying capacity, K = 1000 each (under mutation-drift equilibrium K = N_e). All adjacent marine populations exchanged 100 (TSS) or ten (NSS) migrants per generation (M, symmetrical gene flow) with some long-distance migration also allowed between every second population at a rate of ten times less than that between adjacent populations (Supplementary Fig. 3). To allow frequencies of freshwater adapted alleles to build up and filter to the marine populations as standing genetic variation, a refuge freshwater population (K = 10000) with high frequencies of freshwater adapted alleles was allowed to exchange migrants with the two most central of the ten marine populations (symmetric gene flow with rate M = 1; Supplementary Fig. 3). After a burn-in, two focal freshwater populations were founded from the two marine populations situated at the opposite ends of the stepping-stone chain, after which symmetrical gene flow was allowed between the focal freshwater populations and their closest marine populations at rates M = 1 for TSS and M = 0.2 for NSS (Supplementary Fig. 3). Thus, the freshwater populations were founded from different marine populations with ancestral freshwater adapted alleles stemming from the same refuge freshwater population situated equally far from the two focal freshwater populations, in agreement with the transporter hypothesis. The simulations were then run for 5000 generations to simulate post glacial colonisation of freshwater habitats (10000 years ago, assuming a generation time of two; Baker 1994; DeFaveri and Merilä 2013; DeFaveri et al. 2014). The above parameters generated patterns of IBD and population structuring similar to what was estimated from the empirical data (see Results) and are also consistent with previous studies (DeFaveri et al. 2014).

Three different genetic architectures of a single trait coded by five (independent) bi-allelic loci under stabilising selection for different optima were simulated in the marine (optimal phenotypic value = 0) and freshwater populations (optimal phenotypic value = 20). In architecture A, one large effect additive locus with the homozygote for allele 1 (“a1”) yielding a genotypic value of zero (locally adapted to marine habitat) was simulated; the same was done for the homozygote for allele 2 (“a2”) yielding a genotypic value of 20 (locally adapted for the freshwater habitat). The four remaining minor effect additive QTL were given the allelic value of zero for a1 and allelic values [1,1,2,3] for a2. An empirical example of this kind of architecture is the Ectodysplasin (EDA) gene in the three-spined stickleback (Colosimo et al. 2005). In architecture B, the allelic values of a2 alleles were [1,2,3,4,6] and the allelic values of the alternative alleles were zero, i.e. a single large effect locus is lacking. In architecture C, the distribution of allelic values was the same as in architecture A, but with a2 being recessive to a1. An empirical example of this is the Pitx1 locus controlling pelvic reduction in three- and nine-spined sticklebacks (Cresko et al. 2004; Shapiro et al. 2004).

We simulated ten chromosomes of 100 centi Morgan (cM) with 100 neutral loci (and the QTL) equally spaced across the genome (i.e. uniform recombination rate across chromosomes, with no tight linkage between any two loci). Mutation rate (μ) was set to 10⁻⁸ for all loci, and simulations were initiated with allele frequencies for all loci set to 0.5. At the end of the simulations (generation 15000) we estimated allele frequency differences as the fixation index, F_ST (Weir and Cocherham 1984), with the “stamppFst” function in the R-package “StAMPP” (Pembleton et al. 2013). Further details of parameter settings and the simulation code can be found in DRYAD (XXX).

We classified loci being involved in parallel evolution if the F_ST between the focal freshwater populations (pooled) and all marine populations (pooled) was higher than 0.5 (marine-freshwater F_ST) and lower than 0.5 between the two focal freshwater populations (freshwater-freshwater F_ST). Any loci with both high marine-freshwater F_ST and freshwater-freshwater F_ST (above 0.5) were classified as being locally adapted, but not involved in parallel evolution.

Results

Heterogeneity of pelvic reduction in the wild

Re-analysis of previously published phenotypic data from the wild confirmed a high degree of variation among different populations with respect to pelvic spine and girdle lengths and their inter-correlations (Fig. 1, Supplementary Table 1 and Supplementary Fig. 2). For instance, while spines were absent and pelvic girdles strongly reduced (but not completely absent) in RYT, spines and girdles were completely lacking in the MAS population (Fig. 1 and Supplementary Table 1). Furthermore, in the BYN population spines were absent but pelvic girdles were only partially reduced, whereas in BOL (a population adjacent to MAS), large variation in both spine (SD = 0.037) and girdle (SD = 0.047) lengths was observed, although these two traits were strongly correlated (r² = 0.61, Fig. 1, Supplementary Table 1 and Supplementary Fig. 2). This suggests that a large effect locus affecting both pelvic spines and girdles segregated in this population at the time of sampling. In six pond populations (viz. ABB, KAR, KRK, NAV, HAN, and PYÖ; Fig. 1), relative spine (mean = 0.079) and girdle (mean = 0.15) lengths were only slightly smaller relative to the marine populations (0.11 and 0.16 for spine and girdle lengths, respectively; Supplementary Table 1). Lack of pelvic reduction (relative to marine populations) was observed only in one pond population (KAR; Fig. 1, Supplementary Table 1 and Supplementary Fig. 2).

QTL mapping of pelvic reduction in Helsinki × Rytilampi cross

After LD-network based complexity reduction, all QTL mapping analyses were performed on 241 PCs (six sex-linked PCs were removed). Re-analyses of the 279 F₂ progeny from the HEL × RYT cross confirmed a single QTL region on LG7 for both pelvic spine and girdle lengths in the single-locus analyses (Fig. 2a, b and Table 1). In the multi-locus analyses (absolute trait values), LG7 explained 74–86% of the PVE for both spine and girdle lengths, while all other chromosomes individually explained less than 3% of the phenotypic variance (Table 2). An approximately equal amount of the phenotypic variance was explained by alleles inherited from the F₀ male (pond individual) and the F₀ female (marine individual; ∼30% of the total variance for all traits; Table 1), respectively, with 15–21% of the phenotypic variance also explained by dominance effects. This is the expected outcome for a recessive QTL when the F₀ individuals are fixed for different large effect causal alleles, and when no additional smaller effect loci affect the trait (Klug and Cummings 2018).

Table 1. Summary of QTL-mapping analyses.

Each row corresponds to a significant (arbitrarily numbered) QTL region, with the proportion variance explained (PVE), jointly estimated for all PCs (one for each significant LD-cluster) extracted from such regions. Coding indicates whether the QTL was significant for the alleles inherited from the F₁ female (♀), the F₁ male (♂) or the dominance effect (dom). Effect size (β) is based on the first PC extracted from all SNPs from each significant QTL region. “Std.” indicates whether the trait was standardised (Yes) or not (No). P and P_COR represent nominal and corrected P-values from single-mapping four-way analyses, respectively. Results for standardised trait values are only shown for QTL that were not also significant for absolute trait values.

Table 2. Proportion phenotypic variance explained (PVE) in pelvic traits.

Shown are the percentages of total phenotypic variance explained by different linkage groups (LG), by all SNPs (Tot), by loci inherited from females (♀) and males (♂), as well as the dominance effect (Dom). Shown are results for each cross and trait separately and for absolute trait values.

Figure 2. Quantitative trait locus mapping of pelvic reduction in three independent stickleback crosses.

Single-mapping four-way analyses of four morphological traits associated with pelvic reduction in (a-b) HEL × RYT cross, (c-d) HEL × BYN cross and (e-f) HEL × PYÖ cross. Shown are QTL for pelvic spine length and girdle length with x-axis indicating position in centi Morgans (cM). Results are based on permutation and dotted vertical line indicates genome wide significance at α = 0.05. Results are shown separately for alleles inherited from the male F₁ (M), the female F₁ (F) together with the dominance effect (D) according to legend. Absence of a dominance effect indicates that the trait inheritance is additive, whereas a peak only for M or F indicates that the allelic effect was segregating in the F₀ male or the F₀ female, respectively (see Supplementary File S1 for details). Candidate genes involved in pelvic development are indicated with black text representing genes that affect expression of Pitx1 and red text indicating genes that affect pelvic development downstream of Pitx1 expression. Results for analyses based on relative trait values can be found in Supplementary Figure 4.

QTL mapping of pelvic reduction in Helsinki × Bynästjärnen cross

Among the 308 F₂ progeny of the HEL × BYN cross, single-locus mapping analyses on pelvic spine lengths detected two significant QTL on LG15 (PVE = 8.9%) and LG16 (PVE = 13.7%; Fig. 2c and Table 1) for alleles deriving from the F₀ male (pond individual). Thus, the causal alleles for these QTL segregated in the F₀ pond male (as explained in Supplementary File 1). No dominance effects were detected for these QTL, suggesting that the allelic effects were additive. One QTL on LG6 with an allelic effect deriving only from the F₀ female was also detected (Fig. 2c and Table 1). The QTL significance profiles, in particular for LG15 and LG16, spanned large genomic regions with no obvious peaks (in contrast to LG7 in the HEL × RYT cross; Fig. 2). This remained true when analysing all SNPs individually (fine mapping; Supplementary Fig. 4). The individual and multi-locus phenotypic effects on spine lengths for the QTL on LGs 6, 15 and 16 (using the most significant QTL for each QTL region) are detailed in Figure 3. In the absence of any large effect loci, and when all QTL are additive and independent (as is the case here), phenotypes in the F₂ generation are expected to be normally distributed (Klug and Cummings 2018). For some multi-locus genotypes (of the QTL on LGs 6, 15 and 16), the distribution of spine lengths were approximately normally distributed, except with long tails of highly reduced spine lengths, implicating that some of these loci could be involved in epistatic interactions (this was investigated further in Supplementary File 3). One significant male QTL on LG4 for girdle length was also detected (Fig. 2d and Table 1). Results for relative trait values were highly similar to the absolute trait values, except for an additional significant male and female QTL on LG1, as well as another significant female QTL on LG16 (Supplementary Fig. 5, Table 1 and Supplementary Table 3).

Figure 3. Epistatic interactions in pelvic spine development for HEL × BYN cross.

(a) individual genotype effects where the genotype is given by the first PC (scaled between 0 and 1) from the cluster of SNPs that was the most significant for spine length on LG6, LG15 and LG16 (see Fig. 2), respectively. Genotypes were further based on the genotypes [x_dij, x_sij] depending on which of these were significant for the QTL effects (x_dij for LG6 and x_sij for LG15 and LG16). Some individuals have genotypes between 0 and 1 only because the genotype is based on the first PC of large sets of highly but not perfectly correlated SNPs. Slope of the regression line (grey) is shown. (b) distribution of spine lengths for all multi-locus genotypes from (a). The multi-locus genotype was based on rounding PC1 coordinates from (a) where values below 0.5 (left of vertical dashed line) were considered as “Allele 0” or below or equal to 0.5 were considered as “Allele 1”. The first digit for genotypes in (b) represents thus the genotype of the QTL on LG6, followed by LG15 and LG16, respectively.

Multi-locus analyses (for absolute trait values) identified 11 LGs that accounted for at least 3% of the phenotypic variance in pelvic spine or girdle lengths in the HEL × BYN cross (Table 2). The largest of these effects were found on LG15 and LG16, which explained 9% and 12% of variation in pelvic spine lengths, respectively (Table 2). For pelvic spine and girdle lengths, 39% and 32% of the PVE, respectively, were accounted for by all SNPs in the dataset, which equates to the narrow sense heritability (h²) also accounting for dominance (but not epistatic interaction) effects. For spine lengths, 24% of the total PVE was attributed to alleles deriving from the F₀ male and 17% was attributed to alleles deriving from the F₀ female; only 1% was attributed to the dominance effect (Table 2). For girdle lengths, the corresponding numbers were 16%, 6% and 13% (Table 2).

QTL mapping of pelvic reduction in Helsinki × Pyöreälampi cross

In the HEL × PYÖ cross (283 F₂ individuals), one QTL for spine length was found on LG9, which was explained by alleles inherited from the F₀ male. Two significant QTL for girdle length were found, one on LG19 (explained by alleles inherited from the F₀ male) and the other on LG4 (explained by alleles inherited from the F₀ female; Fig. 2e, f, and Table 1). In multi-locus analyses, the PVE for different LGs mirrored these results; the LGs that contain significant QTL explain most of the PVE, while PVE for all other LGs were <4 % (Table 2). In multi-locus analyses, PVE for pelvic traits was lower than in the HEL × BYN cross; 14% and 16% for spine length and girdle length, respectively, with <2% PVE due to dominance effects (Table 2). When analysing relative trait values, one additional female QTL peak was found for both girdle and spine lengths on LG1 (Supplementary Fig. 5, Table 1 and Supplementary Table 3). Due to the large size of the identified QTL regions, it is not possible to know whether this QTL and that on LG1 detected in the HEL × BYN cross are due to the same or different underlying causal mutations (we consider this as a single QTL region).

Trait correlations in the QTL crosses

There was a strong correlation between relative pelvic spine and girdle lengths (r² = 0.85, Fig. 1, Supplementary Fig. 2 and Supplementary Table 1) in the HEL × RYT cross as expected, since pelvic reduction in this cross is controlled by a single large effect QTL affecting both traits. However, in the HEL × BYN cross, correlation between relative pelvic spine and girdle lengths was much weaker (r² = 0.11, Fig. 1, Supplementary Fig. 2 and Supplementary Table 1). This finding is consistent with pelvic spine and girdle reductions being independently controlled by different QTL. Furthermore, of the 306 F₂ individuals, only four displayed complete lack of spines, despite the fact that the BYN population is fixed for complete spine reduction in the wild (and the spine was absent in the F₀ male). This is consistent with spine length being controlled by multiple additive loci in the HEL × BYN cross. Among the F₂ individuals from the HEL × BYN cross, relative girdle lengths were normally distributed with much smaller variances (SD = 0.012) compared to the HEL × RYT (SD = 0.041) with only two individuals with reduced girdles (Fig. 1, Supplementary Fig. 2 and Supplementary Table 1). This is consistent with the lower heritability of pelvic girdle lengths in the HEL × BYN than in the HEL × RYT cross, with contributions from many small effect loci. In the HEL × PYÖ cross, phenotypic variation was comparable to that in the HEL × BYN cross (SD = 0.012), although the mean relative spine length was slightly smaller (0.078 vs 0.091; Supplementary Table 1) and the mean for relative girdle length was slightly larger (0.175 vs 0.169; Supplementary Fig. 2; Supplementary Table 1).

Scanning for Pel deletions in the full-genome sequence data

In the full-genome re-sequencing data of wild collected individuals, a large deletion upstream of Pitx1 was fixed for all 21 individuals from Rytilampi, where pelvic reduction maps to this region (Fig. 4). The deletion is around 3.5 kb in size and fully encompasses the Pel-2.5kb^SALR region of the three-spined stickleback (Chan et al. 2010). No comparable deletion was found in any other individuals in the dataset (Fig. 4b). This included the two White Sea populations where either complete reduction of both the pelvic spines and girdles was observed (MAS), or a putative large effect locus affecting both spine and girdle length was found to be segregating (BOL; Fig. 1; Supplementary Fig. 2 and Supplementary Table 1).

Figure 4. Sequence depths for H2afy-Pitx1 region.

(a) Relative sequencing depth across the H2afy-Pitx1 intergenic region for 10 individuals from five different populations. Blue line and shading indicate population average and 95% confidence intervals, respectively. The 3.5kb deletion in RYT, seen as a dip sequencing depth, fully encloses the Pel-2.5kb^SALR of three-spined stickleback indicated by the red box. (b) Normalised sequencing depths for the Pel-2.5kb^SALR region for 563 samples from 27 populations show that the complete deletion of the Pel region is unique to RYT. Phenotypic data in the Russian populations MAS and BOL suggest a large effect locus is responsible for pelvic reductions, and BYN and PYÖ are populations analysed here where pelvic reduction does not map to LG7.

Candidate genes

Seven candidate genes or regulatory elements for pelvic reduction from the literature (Supplementary Table 4) were found in LGs with significant QTL (Fig. 2). Hif1a (Mudie et al. 2014), Pel (Chan et al. 2010) and Pou1f1 (Kelberman et al. 2009) are known to regulate the expression of Pitx1, whereas four genes (Fgf8, Wnt8c, Wnt8b and Hoxd9) are involved in the pelvic fin/hind limb development downstream of Pitx1 (Don et al. 2012, Tanaka et al. 2005). However, aside from Pel (LG7) only three of these, Wnt8c (LG6), Hif1a (LG15) and Pouf1f1 (LG16) were clearly within the significant QTL regions (Fig. 2). One candidate locus, Hif1a, is also on LG1 where significant QTL peaks were found when analysing relative (but not absolute) trait values (Table 1 and Supplementary Fig. 5).

Proof of concept simulations

There were no trends in QTL allele frequencies in the simulated populations prior to local adaptation, showing that the 10000 generation burn-in was adequate (Supplementary Fig. 6). In the empirical data from the marine populations, the slope of the regression line between linearised F_ST and geographic distance (point distance in km) was 2.13^-7 (95% of the bootstrap replicates between 2.05^-7 and 2.2^-7) and 1.53^-8 (95% of the bootstrap replicates between 1.10⁻⁸ and 1.96^-8) for nine- and three-spined sticklebacks, respectively. Thus, the slope of the IBD regression line in this dataset was 13.9 times higher for nine-spined sticklebacks compared to three-spined sticklebacks (Fig. 5a). When scaling the geographic distance in the simulated data to the empirical data, the distance between the two simulated marine populations furthest away from each other (i.e. the populations from which the focal freshwater populations were founded) equalled 264 km and 352 km for three- and nine-spined sticklebacks, respectively (Fig. 5a). Thus, with comparable levels of IBD as in the empirical data, our simulations mimic the levels of parallel evolution that can be expected in three- and nine-spined sticklebacks at relatively short geographic scales (<400km). The difference in IBD between the two species is also close to that in the empirical data (264/352 = 0.75). In the empirical data, genetic differentiation between freshwater habitats for populations <400 km from each other (mean F_ST = 0.19 and 0.49 for three- and nine-spined sticklebacks, respectively) was also on par with the simulations (mean [across simulation replicates] F_ST = 0.21 and 0.58 for three- and nine-spined sticklebacks, respectively). Notably, F_ST was >0.8 (linearised F_ST>4) between Pyöreälampi (no pelvic reduction) and Rytilampi (pelvic reduction controlled by Pitx1/Pel), although these ponds are situated only 15 km from each other (Fig. 5a; max and median F_ST = 0.96 and 0.62, respectively, for all pairwise freshwater-freshwater comparisons). This is higher than the F_ST between any pair of three-spined stickleback populations in the data (max = 0.78 and median = 0.26 for all pairwise freshwater-freshwater comparisons). Note that the benchmark empirical data stem from Northern European populations and may thus not accurately represent interspecific differences in population demographic parameters elsewhere.

Figure 5. Simulation results.

Linearised F_ST against geographic distance (IBD) for empirical and simulated data (a), with slope and squared Pearson’s product moment correlation coefficient indicated. Geographic distance for simulated data (IBD in the sea) is scaled to match the slope for the IBD-plot in the sea in empirical data. Freshwater-freshwater F_ST against marine-freshwater F_ST from all QTL from all simulated data (n = 100), with effect sizes indicated as shown in legend (b). Loci in upper left quadrant are classified as being involved in parallel evolution, and loci in the upper right quadrant are loci that are involved in local adaptation in only one freshwater population. This data is summarised in (c) focusing on the four largest effect loci, with genetic architecture and effect sizes indicated by the figure titles.

Do population demographic parameters influence local adaptation?

In the simulations, the relationship between marine-freshwater F_ST and freshwater-freshwater F_ST depended on both the species and genetic architecture (Fig. 5b). For instance, when the genetic architecture included one additive large effect locus (architecture A), this locus was often involved in parallel evolution in three-spined sticklebacks (65% of replicates), but not in nine-spined sticklebacks (3% of the replicates). In 20% of the replicates for both species, the freshwater allele for this locus was fixed in only one of the focal freshwater populations (i.e. local adaptation, but not parallel evolution). When the trait under selection was controlled by several medium effect loci (architecture B), parallel evolution was more common in both stickleback species, particularly for the locus with the largest allelic value (20% and 80% for nine- and three-spined sticklebacks, respectively; Fig. 5b, c). Local adaptation also occurred in 57% of the replicates in nine-spined sticklebacks, and in 18% in the three-spined sticklebacks; Fig. 5b, c). For the locus with the second highest allelic value, the cases of both parallel evolution and local adaptation collectively dropped to 38% and 43% for three- and nine-spined sticklebacks, respectively. With one non-additive large effect locus with recessive alleles locally adapted to freshwater (architecture C), this locus was less likely to be involved in parallel evolution in the three-spined sticklebacks (48%), compared to architecture A. However, results for the nine-spined sticklebacks were similar to architecture A (6% parallel evolution; Fig. 5b and c). Thus, parallel evolution is an expected outcome in three-but not nine-spined sticklebacks, at relatively short (<400 km) geographical distances, particularly when a single additive large effect locus is responsible for freshwater adaptation.

Does local adaptation depend on ancestral allele frequency?

In the simulations, the frequency of a2 allele (locally adapted to freshwater) in each freshwater population after 2000 generations post-colonisation from the sea was always correlated with the a2 allele frequency in the founding marine population, demonstrating that local adaptation depended on ancestral genetic variation (Supplementary Fig. 7). This dependency was the strongest for small effect loci (phenotypic effect of a2 = 1). In these cases, the linear regression slope (β) ranged between 1 and 1.4, the y-intercept ranged between 0 and 0.09, and Pearson’s squared correlation coefficient (r²) ranged between 0.33 and 0.52, and was lowest in three-spined sticklebacks for genetic architecture A (y-intercept = 0.51; β = 4.2; r² = 0.07; Supplementary Fig. 7). In the latter case, standing genetic variation (defined as a2 frequency > 0) was available in 89% of the simulation replicates (considering each freshwater population independently), of which 76% resulted in local adaptation (defined as a2 frequency > 0.5) in at least one of the freshwater populations. The corresponding numbers for nine-spined sticklebacks were 41% and 39%, respectively. Here the correlation between ancestral a2 frequency and a2 frequency post colonisation was also stronger than for three-spined sticklebacks (y-intercept = 0.04; β = 10.8; r² = 0.33), indicating a stronger dependence between ancestral variation and local adaptation. When standing genetic variation was available for local adaptation, three factors likely contributed to lower levels of local adaptation and parallel evolution in nine-spined sticklebacks: i) the smaller K in nine-spined stickleback freshwater populations (500 vs. 1000), which is expected to result in faster genetic drift; ii) lower (1/5th) post-colonisation gene-flow between adjacent marine and freshwater populations; and iii) lower a2 frequency (when present as standing genetic variation) in nine-compared to three-spined sticklebacks (2.9% vs. 3.4%). Thus, IBD in the sea likely limits the probability that a freshwater adapted allele is maintained as standing genetic variation, while smaller effective population sizes and reduced gene flow can further limit local adaptation, even when ancestral variation in the founding population is present (in low frequencies). The generally higher β’s (3.20 vs. 2.15) and y-intercepts (0.11 vs. 0.05) for QTL with a2<6 for nine-spined sticklebacks indicate that smaller effect QTL mostly have a stronger influence on local adaptation in nine-compared to three-spined sticklebacks (Supplementary Fig. 7), resulting in polygenic and/or incomplete local adaptation (Supplementary Fig. 6 and Supplementary Fig. 8).

Discussion

This study demonstrates that the genetic basis of pelvic reduction in nine-spined sticklebacks is more variable than in three-spined sticklebacks. This is consistent with the empirical evidence that IBD in the sea is stronger, and that freshwater populations are more isolated in nine-than three-spined sticklebacks. According to our simulations, such scenario would limit the pool of ancestral variation available for freshwater adaptation in nine-spined sticklebacks compared to three-spined sticklebacks. Thus, the level of ancestral variation available for local adaptation (and also for parallel evolution) is likely a function of population demographic parameters, with local adaptation being less likely in poorly connected species, as suggested by Merilä (2013, 2014). However, other non-mutually exclusive factors, such as the genetic architecture (i.e. dominance effects, heritability, mutation rates and numbers of causal loci involved), as well as non-parallelism in phenotypic selection optima are also likely to play roles. In the following, we discuss the possible causes of the discrepancy in pelvic structure development between nine- and three-spined sticklebacks, as well as their implications to our understanding of adaptive evolution in the wild.

Can the genetic architecture of pelvic reduction be explained by population demographic history?

Together with earlier QTL studies (Shapiro et al. 2006, 2009; Shikano et al. 2013), we show that multiple genomic regions (11 QTL, of which ten are novel to this study) are associated with pelvic reduction in nine-spined sticklebacks across their distribution range. Only one small effect QTL region (LG1; Table 1 and Supplementary Fig. 5) was shared between any two crosses (HEL × BYN and HEL × PYÖ), but even here it is not certain that the underlying causal mutations are the same. Support for genetic parallelism among marine-freshwater pairs of three-spined stickleback populations is strong (Colosimo et al. 2005; Jones et al. 2012; Terekhanova et al. 2014, 2019; Nelson and Cresko 2018). Although the high frequency of Pel deletions (disrupting pelvic armour development) means that most of the deletions associated with pelvic reduction in three-spined sticklebacks are independently derived (Xie et al. 2019), Pitx1/Pel is predominantly responsible for pelvic reduction in three-spined sticklebacks (Chan et al. 2010; Xie et al. 2019). On the other hand, almost all of the studied nine-spined stickleback populations to date display different QTL regions responsible for pelvic reduction (of varying QTL effect sizes and dominance relationships). One explanation for this could be that the independently derived Pel deletions (among other possible solutions for pelvic reduction) are usually the most favoured by selection, and when they are accessible as standing genetic variation, they quickly spread and become fixed in newly colonised freshwater populations. However, with a more limited pool of ancestral variation, freshwater adaptation in nine-spined sticklebacks could be restricted to less optimal solutions. This is consistent with our results that showed only marginal, partial or no pelvic reduction in seven out of the ten pond populations, and exemplified by the HEL × PYÖ and HEL × BYN crosses where pelvic reduction was polygenic and much less heritable than in the HEL × RYT cross. This hypothesis predicts that if the pelvic reducing Pel-deletion from Rytilampi (or any other allele with comparable fitness effects) would invade pond populations that currently do not display complete pelvic reduction, this mutation would quickly spread and become fixed. Such rapid spread of freshwater adapted EDA alleles has been well documented in the three-spined stickleback (Schluter et al. 2010; Bell et al. 2004).

Using simulations with population demographic parameters similar to those estimated from empirical data, we demonstrated that parallel evolution is more likely to occur in three-than in nine-spined sticklebacks – that is, when independently founded freshwater populations have access to the same ancestral standing genetic variation. This is consistent with the large body of studies showing that parallel evolution among pairs of marine-freshwater ecotypes in three-spined sticklebacks is commonly due to the repeated use of the same ancestral variation from a global pool of standing genetic variation in the sea, both at the global (Jones et al. 2012) and more local scales (Colosimo et al. 2005; Jones et al. 2012; Terekhanova et al. 2014, 2019; Nelson and Cresko, 2018). Furthermore, while pelvic reduction in freshwater populations of three-spined stickleback populations is more or less the norm (Colosimo 2005; Klepaker et al. 2013), our survey of phenotypic data from the wild demonstrates that full pelvic reduction in northern European freshwater populations of nine-spined sticklebacks is rare. In fact, it only occurs in two of the ten populations, with one additional population segregating for a large effect locus and one showing reduction in spines but not in girdles. Consistent with these results, when the access to ancestral standing genetic variation was restricted in the simulations (by IBD in the sea and population structuring among freshwater populations being similar to that observed in nine-spined sticklebacks), parallel evolution was much less likely, and local adaptation was less complete or lacking altogether. Furthermore, (partial) local adaptation was often polygenic and due to fixation of phylogenetically independent (small effect) QTL, and not due to parallel evolution.

In our simulations, genetic parallelism in the three-spined stickleback was less common when the large effect freshwater adapted allele was recessive (architecture C), compared to when it was additive (architecture A). This may also partly explain why the Pitx1/Pel deletions responsible for pelvic reduction in three-spined sticklebacks are mostly due to repeated independent deletions rather than phylogenetically related alleles (Xie et al. 2019). In contrast, the additive EDA gene controlling lateral armour plate numbers in three-spined sticklebacks is a classic example of local adaptation from ancestral standing variation (Colosimo et al., 2005; Schluter and Conte 2009). Furthermore, if a single large effect QTL already leads to a near-optimal phenotype, any additional minor or medium effect loci could lead individuals farther from the optimum, i.e. maladaptation (Bolnick et al. 2018). If so, selection is expected to prune such variants, leaving only large effect QTL in the population. This could explain why other small or medium effect loci for pelvic reduction are rare – but not completely absent – in three-spined sticklebacks as well as in Rytilampi nine-spined sticklebacks. Consistent with this, the single large effect locus available for freshwater adaptation (architectures A and C) was the most commonly used allele for local adaptation at the expense of medium effect loci; using both would have led to overshooting the phenotypic optimum (in both species; Supplementary Fig. 8; Rogers et al. 2012).

Geographic heterogeneity in selection optima

Another explanation for the high heterogeneity in the genetic architecture of pelvic reduction in the nine-spined stickleback resides within habitat environmental variation resulting in different selection optima in different pond populations (cf. Stuart et al. 2017; Thompson et al. 2019). The small differences between pelvic reduced phenotypes in our study (e.g. in BYN and MAS/BOL spines are completely absent, while in RYT they are only strongly reduced; Fig. 1, Supplementary Table 1 and Supplementary Fig. 2) could in fact indicate different selection optima in the different populations (Stuart et al. 2017; Thompson et al. 2019). It is possible to use available phenotypic data to estimate the phenotypic optima of pelvic morphology (hypersphere) in each of the populations using Fisher’s geometric model (Stuart et al. 2017; Thompson et al. 2019), where a strong overlap would suggest a higher probability of genetic parallelism (Thompson et al. 2019). However, this assumes that the populations have access to exactly the same ancestral variation and are free to evolve and reach their optima, which is at odds with the results presented here. Without detailed environmental data or direct estimates of strength of selection on pelvic phenotypes, disentangling the effects of gene flow and within habitat environmental variation (assuming this leads to non-parallel angles of selection) is not possible (Stuart et al. 2017).

In a recent simulation study, Thompson et al. (2019) showed that genetic parallelism from standing genetic variation rapidly declines as selection starts to change from fully parallel (optima angle of 0°) to divergent (optima angle of 180°), especially when the trait is polygenic. However, although selection was fully parallel in our simulations, we failed to observe strong genetic parallelism for smaller effect loci (with allelic effects < 6) in both the three- and nine-spined stickleback-like scenarios (Fig. 5a, c). This suggests that the effects of the underlying genetic architecture on parallelism (in conjunction with some IBD and population structuring) can be independent of the angle of optimal phenotypes between two habitats. In other words, non-parallelism in selection optima is not necessary to explain non-parallelism at the phenotypic and genetic levels. Thus, it is important not to disregard the population demographic setting as a factor that could severely restrict heritability for adaptation and/or constrain adaptation to less optimal solutions. Evolutionary studies of species with population demographic parameters comparable to those typical for vulnerable or endangered species/populations (such as the nine-spined stickleback) would be valuable to gain a better understanding of how such species may respond to environmental changes and urbanisation (Thompson et al. 2018).

The limited role of Pitx1/Pel in the nine-spined stickleback pelvic reduction

Thus far, Pitx1 has been found to be responsible for pelvic reduction in nine-spined sticklebacks in both Canada (Shapiro et al. 2006) and Finland (Shikano et al. 2013; this study), but we do not know to what extent the underlying mutations are identical by descent. By scanning whole-genome sequence data of 27 populations, we found a deletion in the genomic region expected to contain Pel (LG7) only in Rytilampi – where pelvic reduction maps to this gene. Similar to marine-freshwater pairs of three-spined sticklebacks, we would have expected Pitx1/Pel deletions to exist in several other nine-spined stickleback pond populations as well, most notably in the MAS and BOL populations, where phenotypic data suggest large effect QTL to be responsible for pelvic reduction. Since no deletions were found in the genomic region spanning Pel and Pitx1 in MAS and BOL, Pitx1/Pel is likely not responsible for the pelvic reduction in these two populations. This provides further evidence that the genomic basis of pelvic reduction in the nine-spined stickleback is more heterogeneous than that in the three-spined stickleback. The loci, most likely major effect, responsible for pelvic reduction in these ponds are yet to be identified.

Pelvic reduction outside marine-freshwater study systems

While the evidence for genetic parallelism at large geographical scales in the marine-freshwater stickleback model system is extensive (Colosimo et al. 2005; Jones et al. 2012; Terekhanova et al. 2014, 2019; Nelson and Cresko 2018; Fang et al. 2019), the level of parallelism in lake-stream and pelagic-benthic ecotype pairs of three-spined sticklebacks is much more diverse (Peichel et al. 2001, 2017; Conte et al. 2012, 2015; Stuart et al. 2017). For instance, Conte et al. (2015) found that among benthic-limnetic three-spined stickleback pairs from Paxton and Priest lakes (Vancouver Island, BC, Canada), 76% of 42 phenotypic traits diverged in the same direction, whereas only 49% of the underlying QTL evolved in parallel in both lakes. For highly parallel traits in two other pairs of benthic-limnetic sticklebacks, only 32% of the underlying QTL were shared (Conte et al., 2012). Similarly, Stuart et al. (2017) found that among 11 independent evolutionary replicate pairs of lake-stream three-spined stickleback populations (Vancouver Island, BC, Canada), both within habitat variation and constraints to gene flow contributed to the observed variation in levels of phenotypic parallelism. Different lakes and streams do not likely have similar access to the same global pool of ancestral variation as pairs of marine-freshwater three-spined sticklebacks, where gene flow in the sea is high. This is consistent with the notion that more heterogeneous access to ancestral variation can indeed limit genetic parallelism. This is also true for one example of marine-freshwater three-spined stickleback divergence among isolated insular freshwater populations in the Haida Gwaii archipelago off the northern Pacific coast of Canada (Deagle et al. 2013). Here, similar to the nine-spined sticklebacks in this study, several freshwater populations did not display any reduction in pelvic armour. However, those populations that were fully plated were also genetically more similar to adjacent marine individuals, suggesting that recent marine-freshwater admixture and/or selection favouring plated freshwater individuals could explain this pattern. Other morphological traits, such as lateral plating (coded by the EDA locus) were instead more consistent with the same global variants being repeatedly re-used at the regional scale. Thus, with respect to access to ancestral variation available for freshwater adaptation, nine-spined sticklebacks are likely closer to the three-spined stickleback lake-stream and benthic-limnetic study systems than to the three-spined stickleback marine-freshwater study system. The only notable exception is the lake-stream three-spined sticklebacks mentioned above, where genetic structuring also is high.

Epistatic control of pelvic reduction?

For traits with more than one additive QTL of equal effect sizes, the F₂ phenotypes are expected to be normally distributed with a mean close that of the mean for the parents (Klug and Cummings 2018). This was not the case for spine length in the HEL × BYN cross, which was controlled by three QTL with similar effect sizes. In this case, the bulk of the phenotypic values was centered around the mean, but also had a long tail of individuals with strongly reduced pelvic spines (Fig. 1, Fig. 3 and Supplementary Fig. 2). This skew in the phenotype distribution could be caused by epistatic interactions among loci controlling pelvic spine length (Wolf et al. 2000). Consistent with this hypothesis, complete spine reduction was most likely when the allele responsible for spine reduction for the LG6 QTL was combined with at least one allele causing spine reduction from the LG15 and LG16 QTL (Fig. 3b and Supplementary File 3). If a threshold number of alleles are needed for complete pelvic reduction, this could also explain how standing genetic variation in the sea is maintained, as the necessary multi-locus genotypes that cause sub-optimal phenotypes in the sea are rarely formed, due to overall lower frequencies of spine-reducing alleles in the sea. This is analogous to “epistatic shielding” that can contribute to the persistence of disease alleles in populations (Phillips and Johnsson 1998; Phillips 2008). Consistent with this possibility, LG6 of the F₀ female of the HEL × BYN cross (from the sea) was polymorphic for the pelvic spine QTL effect (Fig. 2) – evidently a single pelvic-reducing allele alone in this female was not enough to cause any pelvic reduction at all (this female had a complete pelvis).

Candidate genes

While the QTL peak for the Pitx1/Pel region in the HEL × RYT cross was narrow, this was not the case for the other QTL we detected. Hence, due to the large QTL regions detected by four-way single-mapping analyses, it was not meaningful to perform gene ontology enrichment (GO) analyses – the QTL regions would have contained possibly thousands of genes. Instead, we searched the literature for known candidate genes for pelvic reduction and found three (excluding Pitx1/Pel) that were clearly contained within the identified QTL regions (Fig. 2 and Supplementary Table 4). Due to the aforementioned large QTL regions, these can only be considered as highly putative candidate genes for pelvic reduction and will not be discussed further. However, further studies of pelvic reduction might find these candidates worthy of attention.

Conclusions

Our results show that the repeated parallel reduction in pelvic structures in freshwater populations of nine-spined sticklebacks is due to a diverse set of genetic changes: only one small effect QTL for pelvic reduction was shared between the three experimental crosses in this study. In one cross, pelvic reduction was mapped to the previously identified Pitx1/Pel regions, but in the two other crosses, the genetic basis of pelvic reduction was polygenic, and mapped to many different chromosomes. In addition to these, yet another large effect QTL different from the Pitx1/Pel locus likely segregates in one nine-spined stickleback population waiting to be identified. The results also shed light on the possible drivers of the observed genetic heterogeneity underlying pelvic reduction; as shown by simulations, heterogeneous genetic architectures are more likely to emerge when access to ancestral variation is limited by strong isolation by distance and population structuring. This reinforces the role of the nine-spined sticklebacks as a useful model system, alongside the three-spined stickleback, to study adaptive evolution in the wild. Furthermore, since the population demographic characteristics of nine-spined sticklebacks are similar to small and endangered species/populations, it is also likely to be a well-suited model to study genetics of adaptation in populations of conservation concern.

Footnotes

• The results have not changed, but they have been framed slightly differently. The abstract, Introduction and Discussion have been largely re-written. More details have been included into the Material and Methods section.