Epistatic mutations under divergent selection govern phenotypic variation in the crow hybrid zone

Abstract

The evolution of genetic barriers opposing inter-specific gene flow is key to the origin of new species. Drawing from information of over 400 admixed genomes sourced from replicate transects across the European hybrid zone between all-black carrion crows and grey-coated hooded crows, we decipher the interplay between phenotypic divergence and selection at the molecular level. Over 68% of plumage variation was explained by epistasis between the gene NDP and a ~2.8 Mb region on chromosome 18 with suppressed recombination. Both pigmentation loci showed evidence for divergent selection resisting introgression. This study reveals how few, large-effect loci can govern prezygotic isolation and shield phenotypic divergence from gene flow.

Understanding the origin of species has been foundational to the field of evolutionary biology. With the application of genomic approaches in natural populations, uncovering the genetic basis of speciation has come within reach^1,2. All-black carrion crows (Corvus (corone) corone) and grey-coated hooded crows (C. (c.) cornix) meet in a stable and narrow contact zone that has likely formed in the early Holocene^3,4. Assortative mating⁵ and social marginalization of minority phenotypes⁶ based on plumage pigmentation patterns act in concert to reduce the amount of crossbreeding⁷. Genome scans in the European hybrid zone suggest that genetic variation is homogenized genome-wide by inter-specific gene flow beyond the morphologically visible confines of the hybrid zone⁸. A few genomic regions, however, including a ~2.8 Mb region on chromosome 18 with low levels of recombination, have been identified as candidate barrier loci that resist gene flow and might contribute to maintaining phenotypic identity^8,9. A screen of genomic outlier loci at independent contact zones in Asia pointed towards some of the same genes acting in the Wnt signalling component of the melanogenesis pathway³ known to modify pigmentation patterns⁹.

Yet, F_ST outliers do not necessarily reflect the quantitative trait loci underlying phenotypic variation¹⁰. Admixture analyses at contact zones are a powerful means to disassociate population identity from phenotype and firmly establish whether genes coding for phenotypic divergence are indeed subject to divergent selection¹¹. Here, we draw from information of admixed genomes to fine-map the genetic basis of phenotypic divergence and infer the mode and strength of selection modulating contemporary gene flow across the contact zone^11,12. We sampled a total of 409 individuals from replicate transects in the central (N = 173) and southern part of the European hybrid zone (N = 236) (Supplementary Figs. 1 and 2) and further included individuals from allopatric populations to infer ancestry (central N = 75, southern N = 38). All individuals were genotyped for a final set of 1111 SNPs selected from 16.6 million variants segregating in European populations including highly ancestry-informative (high F_ST N = 735; fixed N = 51) and background markers (N = 325) spread across the genome (Supplementary Fig. 3, Supplementary Table 1). Markers were equally spread across the genome except for chromosome 18 that was densely covered with 230 tightly linked markers in the previously identified outlier region⁸ (median linkage disequilibrium (LD) central r² = 0.18, southern r² = 0.17). For a set of downstream analyses, chromosome 18 was treated separately from the remaining genome-wide SNPs (median LD within chromosomes: r² = 0.0024).

To map the genetic basis of phenotypic divergence, we first quantified plumage variation in 129 southern hybrids. We divided the dorsal and ventral plumage surface into discrete patches and scored shading on a grey scale reflecting the amount of eumelanin deposited into the feathering¹³ (Supplementary Fig. 4). Inter-individual variation in shading was highly correlated among patches (average Spearman’s r = 0.76, Supplementary Fig. 5a). Next, we performed PCA, in which 85.88% of the total phenotypic variance was captured by the first two components (PC1: 78.22%, PC2: 7.66%). Positive PC1 scores corresponded to an overall darker appearance (Fig. 1a), whereas PC2 separated the central body parts (belly and mantle) from distal and caudal parts (Supplementary Fig. 5b).

Figure 1. Genetic basis of phenotypic variation.

(a) Examples of pure parental (CC: carrion crow, HC: hooded crow) and hybrid (Fx) phenotypes. Plumage pigmentation was scored on a grey scale for dorsal (D) and ventral (V) patches and subsequently summarized by principal component analysis. The abscissa shows the smoothed density distribution of PC1 scores for 122 individuals with the underlying shading for orientation. (b) Manhattan plot of genome-wide association analysis for PC1 scores shown for 1103 SNPs by chromosome. Significantly associated SNPs are shown in red. The dashed line represents the genome-wide significance threshold. (c) Left: Close-up of the 74 most ancestry informative SNPs within the F_ST outlier region of chromosome 18 (chr18). SNPs are coloured by ancestry of the parental populations (black: carrion crow, light grey: hooded crow, dark grey: heterospecific). Right: Ancestry components inferred across all SNPs group individuals into three discrete classes corresponding to the three diplotypes: chr18_DD, chr18_DL, chr18_LL. (d) Decomposition of phenotypic variation as predicted by the recessive, epistatic interaction between the gene NDP and chr18 ancestry type. Alleles are named by their inferred phenotypic effect (D: dark, L: light).

Genome-wide association mapping uncovered one major genomic region strongly associated with variation in PC1. Out of a total of 170 SNPs remaining significant after multiple testing correction 166 fell within ~2.8 Mb of the genomic outlier region on chromosome 18 (chr18) with single SNPs explaining up to 76.07% (SNP: scaffold_78:1682876) of the variance in PC1 (median: 54.0 %; Fig. 1b, Supplementary Fig. 6, Supplementary Table 2). High linkage disequilibrium among SNPs in this region and low levels of inter-specific recombination allowed categorizing this region by “ancestry diplotypes” (Fig. 1c): a purebred carrion crow type homozygous for the “dark” (D) allele (chr18_DD), a purebred hooded crow type homozygous for the “light” (L) allele (chr18_LL) and the heterospecific type (chr18_DL). These ancestry types predicted segregating phenotypic variation in PC1 equally well as the most significant SNPs (74.29% vs. 76.07%, respectively). Chr18_DD individuals (N = 29) were completely black corresponding to a narrow distribution of high PC1 scores (standard deviation [SD] in PC1 scores = 0.07, range 3.96–4.33) (Fig. 1d). On the contrary, chr18_LL individuals had a much wider phenotypic distribution (SD = 1.46, range -3.74–2.18): while the majority of individuals (N = 35) displayed hooded crow phenotypes, 14 individuals showed intermediate plumage patterns with extensive variation among them. The upper belly and the mantle, however, were grey in all cases. Individuals of the heterospecific chr18_DL type covered almost the entire phenotypic spectrum except for the extremes of pure black or the grey-coated hooded phenotypes (SD in PC1 scores = 1.96, range -3.34–3.59; Fig. 1d).

These results demonstrate that phenotypic variation was largely, but not exclusively, governed by a genetic factor within chr18 containing 88 genes. Additional environmental or genetic factors are required to explain the residual phenotypic variation within chr18_LL and chr18_DL types. Three SNPs in proximity to the genes NDP and EFHC2 on chromosome 1, and a single SNP in the first intron of LRP6 on chromosome 1A explained additional 10.27% and 0.43% of the variance in PC1, respectively (Supplementary Table 2). Norrin (NDP) and low-density lipoprotein receptors LRP5/6 closely interact in the Wnt signalling pathway¹⁴, and had been hypothesized to be involved in phenotypic divergence across multiple independent crow hybrid zones^3,8. Gene expression of NDP had further been associated with pigmentation patterning and divergence between carrion and hooded crows⁹. Recently, NDP has also been suggested to regulate melanin-based patterning in pigeons¹⁵. NDP, rather than EFHC2, is thus a prime candidate to modulate both the intensity and position of plumage pigmentation, consistent with the wide phenotypic distribution in chr18_LL and chr18_DL individuals (Fig. 1d).

Next, we included dominance and epistasis into the statistical model. The best model accounted for 87.91% of the variance in PC1 corresponding to 78.22% × 87.91% = 68.76% of the total phenotypic variance. It supported additive and dominance effects of LRP6, NDP and chr18, as well as recessive epistasis between NDP and chr18 (Supplementary Table 2). Allelic variation in all three SNPs linked to NDP had no phenotypic effect in all-black chr18_DD individuals, but accounted for most of the residual variation in chr18_DL and chr18_LL (Fig. 1d). Individuals of both diplotypes (chr18_DL and chr18_LL) became increasingly lighter in combination with the NDP_DL and NDP_LL genotype. Only chr18_LL individuals homozygous for the light NDP allele (i.e., NDP_LL) recovered the hooded crow phenotype.

All three major loci — chr18, NDP and LRP6 — resided within the few previously identified genomic regions of increased differentiation between parental populations^3,8. Moreover, the 170 significant SNPs identified in the GWAS were among those having the highest F_ST values (GLMM: P-value < 2 × 10^-16; Supplementary Fig. 7). F_ST outliers need not be associated with the genes underlying phenotypic divergence, in particular for polygenic trait architectures¹⁰. The association of all genetic factors controlling phenotypic variation with genetic differentiation among pure, parental populations thus constitutes a first indication that the loci controlling phenotypic divergence may themselves be subject to divergent selection limiting gene flow across the hybrid zone. To quantify selection acting upon admixed genomes segregating in the contemporary hybrid zone, we performed cline analyses. First, we considered gene flow along a spatial axis across the hybrid zone. Transition of genome-wide ancestry across both geographic transects was best explained by a sigmoid cline function (Fig. 2a,e, Supplementary Table 3) supporting previous evidence for recurrent backcrossing³. Inferred cline widths of 316.43 km (central) and 553.85 km (southern) substantially exceed morphologically based estimates of 45–100 km (central) and 10–100 km (southern) (Supplementary Table 4). This provides further evidence for genome-wide introgression expanding far beyond morphologically inferred boundaries^8,16. In contrast to neutral genetic variation, barrier loci under divergent selection that confer a cost to hybrids would fail to move freely across hybrid zones, and are thus expected to show steeper clines and reduced width while maintaining the centre of the cline¹². Consistent with this prediction, chr18 and NDP shared cline centres with genome-wide estimates, but were significantly reduced in width (Fig. 2b,f, Supplementary Fig. 8, Supplementary Table 5). Estimates of 58 km (chr18) and 105 km (NDP) in the central zone closely mimicked morphology-based inference (Supplementary Table 4). Similarly, narrow clines were observed in the southern contact zone for NDP (65 km) and for chr18 (Fig. 2f, Supplementary Fig. 8, Supplementary Table 5), where the formal estimate of 241.35 km was likely inflated due to incomplete sampling of populations with carrion crow ancestry (Fig. 2f).

Figure 2. Cline analysis.

Left: Geographic clines for hybrid indices estimated from genome-wide data (a, e, N = 752 SNPs) and SNPs on chromosome 18 (b, f, N = 230 SNPs). Estimates are shown for both the central hybrid zone (a, b, N = 116 individuals) and the southern hybrid zone (e, f, N = 273 individuals). Depicted are the maximum-likelihood clines and observed average hybrid indices per sample location (with the 95% credible cline region shaded in grey). Each cline extends from the allopatric carrion to the allopatric hooded crow populations. Circle areas reflect sample sizes, colours are used for orientation (dark grey: carrion crow, light grey: hooded crow). The abscissa is centered on the cline centre, rectangles (yellow: central, red: south) depict the cline widths predicted by the models. Right: Summary statistics of genomic cline analyses shown for all 1111 SNPs across the genome (left c, g) and for the outlier region of chromosome 18 (chr18) specifically (right d, h). Depicted is the -log₁₀(P-value) of the cline rate v with high values reflecting departures of introgression from the genome average. SNPs with the strongest evidence (upper 5 percentiles) for reduced introgression are coloured (yellow: central, red: south). The dashed lines in (g) and (h) represent the genome-wide significance threshold.

Genomic cline analysis provides an alternative way to infer recent selection against hybridization. It does not require a spatial axis, but describes locus-specific introgression along a gradient of genome-wide admixture¹⁷. Loci subject to divergent selection are associated with lower fitness in heterospecific genomic backgrounds and are accordingly less likely to introgress into the alternative genomic background. Using the full multi-locus data set, we fitted genomic cline models that describe variation in cline shape with two parameters u and v. Analogous to a shift in cline centre and width in a geographic context, u and v describe the departure of introgression from the genome average. Elevated values of the genomic cline rate v are consistent with underdominance and epistatic incompatibilities, whereas changes in the u parameter are indicative of either directional selection or unequal sampling¹¹. All loci that showed significantly steeper and thus narrower clines (larger v) than the genome average in the south were located in the vicinity of chr18, NDP and LRP6. In the central hybrid zone, patterns were qualitatively similar, but due to the smaller sample size not statistically significant (Fig. 2c,d,g,h, Supplementary Table 6). These results thus complement evidence from the geographic cline analyses supporting divergent selection on all three loci associated with phenotypic variation.

So far, we have treated chr18 as a single locus. This region is characterized by a saddle pattern of F_ST peaks and contains three melanogenesis genes acting upstream of the transcription factor MITF, which is known to play a central role in crow phenotypic divergence^8,9,13 (Supplementary Figs. 9–11). Pinpointing the genes responsible for the large contribution of chr18 to phenotypic variation is impeded by high linkage disequilibrium owing to low recombination, possibly due to an inversion⁸ or proximity to the centromere¹⁸. Nevertheless, AXIN2, located under the first F_ST peak, is a prime candidate. Like NDP, it acts in the Wnt pathway and is involved in pigmentation and patterning¹⁹. PRKCA and a tandem array of CACNG genes act in the MAP-Kinase pathway and reside in the second F_ST peak. This region was also characterized by signatures of positive selection in the lineage leading to the hooded crow (Supplementary Figs. 9h and 10h) and shows the strongest association of all SNPs on chromosome 18 with NDP (P = 2 × 10^-7). While this favours mutations in CACNG/PRKCA as causal in explaining phenotypic variation, a substantially larger sample of recombinant individuals is needed for fine mapping. Our results even raise the possibility that an additional epistatic interaction between AXIN2 and PRKCA/CACNG might be responsible for phenotypic divergence (Supplementary Text).

This study demonstrates the power of leveraging information from naturally occurring hybridization for speciation research¹¹. In combination with outlier scans of parental populations, admixture analyses provide valuable insight into the genetic architecture of phenotypic variation, its genetic control and selection acting upon it. Variation in hybrid colour phenotypes was best attributed to recessive epistasis between two genes predicting the two pure parental forms and three hybrid classes. This corresponds well with previous, morphology-based classification into three distinct hybrid types^20,21. It also adds to the evidence that gene-by-gene interactions are an important component of phenotypic variation and may help explain the missing heritability for quantitative traits²². In contrast to indirect evidence from genome-scans and gene expression analysis from previous work^3,8,9, this study provides conclusive evidence that genes controlling phenotypic variation are subject to divergent selection. There is considerable overlap in the genes identified by genome scans comparing morphologically pure parental populations^3,8 and cline analysis in admixed genomes (this study; Pearson’s correlation between -log₁₀(P-values) of v with F_ST central = 0.71 and south = 0.80). This supports the view that in species that have experienced recent gene exchange the much debated regions of elevated genetic differentiation (“speciation islands”²³) may constitute a valid starting point in the quest for barrier loci relevant to species diversification^1,24. Finally, it is striking that all genomic regions identified to be under divergent selection controlled phenotypic variation resembling findings in a pair of North American parulid warblers²⁵. Considering that divergence in plumage pigmentation patterns reduces gene flow in crows^5,6,26–28, this finding supports the hypothesis that population divergence may start from very few, large-effect loci conveying prezygotic isolation²⁹. It thus strengthens the notion that strong prezygotic barriers may evolve prior to postzygotic isolation³⁰.

Methods

Population sampling

We obtained blood and tissue samples of carrion and hooded crows (Corvus (corone) corone and C. (c.) cornix) and their hybrids along two transects across the European hybrid zone (Supplementary Figs. 1 and 2). Transects were chosen such that they included phenotypically pure populations resembling the parental allopatric populations at the endpoints, and several geographically spaced populations with mixed hybrid phenotypes in between. One transect was located in eastern Germany (and Graz, Austria) and the other at the north-western border of Italy towards France. We hereafter refer to these transects as “central” and “south”, respectively. In the central hybrid zone, we collected 152 nestlings from 76 nests in May–June 2007, 2008, 2013 and 2014. Additionally, we obtained samples from 26 nestlings that were raised in captivity and originated from the central hybrid zone near Graz. Individuals allopatric to the central hybrid zone were caught in north-western Germany (adult carrion crows, N = 45) and Poland and Sweden (adult hooded crows, N = 30). For all these individuals, blood samples were taken from the brachial vein and stored either in Queens’s lysis buffer, EDTA- or heparin-coated sample containers. In the southern hybrid zone, adult birds were shot by local hunters and tissue samples from the breast muscle were stored in Ethanol (N = 238 individuals including 107 samples from ¹⁶ and 131 samples taken by N.S. for this study). Allopatric populations of the southern transect consisted of 26 adult carrion crows from southern Germany and 12 adult hooded crows from central Italy, for all of which blood samples were taken. For more information on samples and sampling locations consult Supplementary Tables 7 and 8.

Permissions for sampling of wild crows were granted by Regierungspräsidium Freiburg (Aktenzeichen: 55-8852.15), Landratsamt Zwickau (364.622-N-Her-1/14), Landratsamt Mittelsachsen (55410704 Beringungserl-Voigt_14), Landratsamt Vogtlandkreis (364.622-2-2-88841/2014), Landratsamt Meißen (672/364.621-Kennzeichnung von Tieren-18935/2013), Landratsamt Bauzen (67.3-364.622:13-01-Krähen), Landesdirektion Sachsen (24-9168.00/2013-4), Landesamt für Verbraucherschutz, Landwirtschaft und Flurneuordnung Brandenburg (23-2347-8a182008) in Germany and by Jordbruksverket (Dnr 30-1326/10) in Sweden. Polish hooded crow nestlings were provided by courtesy of Dr. Andrzej Kruszewicz from the animal rehabilitation centre of Warsaw Zoo. Italian hooded crows of the allopatric population were provided by Centro Recupero Fauna Selvatica LIPU di Roma, Rome, Italy. Crow specimens from the southern hybrid zone were provided by the local administrations (Province di Cuneo ed Alessandria) in the frame of annual spring shooting of bird pests.

Sample preparation and SNP genotyping

We extracted DNA from blood samples using a standard phenol chloroform assay and from muscle tissue using the DNeasy blood and tissue kit (Qiagen). DNA quantity and quality was assessed using Nanodrop and the SYBR green fluorescence assay (Invitrogen).

From a total of 16.6 million variants segregating in the European crow populations^3,8 we selected 1152 SNPs for genotyping with the GoldenGate assay (Illumina). Golden gate primers were primarily designed to contain information on ancestry differentiating between carrion and hooded crows. Using prior knowledge on the degree of differentiation (F_ST) per SNP we primarily chose SNPs with high allelic differential, but also included a set of ancestry-informative makers representing the genomic background (Supplementary Fig. 3). Specifically, the following types of makers were chosen: i) fixed (N = 51) — variants that were fixed between European carrion and hooded crows⁸. ii) outlier (N = 767) — a taxon-informative SNP set differentiating between carrion and hooded crows without fixed differences. This set included SNPs with the highest F_ST values (99^th percentile) residing within localized phylogenetic trees (class I “cacti”) or 50 kb outlier windows with elevated genetic differentiation (95^th percentile of windows-based F_ST; see 3,8). This set also included 4 SNPs within candidate melanogenesis genes that were not part of class I cacti or outlier windows. iii) background (N = 334) — a genome-wide SNP-set that has not been shown to be under selection⁸, but that was still taxon-informative, was defined as variants that fell within the 80–95^th F_ST percentile in regions 100 kb away from outlier regions and not within 10 kb of each other. These included another 14 SNPs residing within candidate melanogenesis genes. Note that due to an extremely right skewed F_ST distribution, the F_ST values of the 80–95^th percentile are low and reflective of the genome-wide average (for F_ST values see below).

In a first round of filtering, we excluded SNPs with a minor allele frequency < 0.1 and SNPs within 100 bp of each other. Forward and reverse primers spanning 60 bp around the target loci were designed with Primer3³¹. In a second filtering step, we excluded SNPs where the primers included blocks of missing sequences or did not map uniquely in the genome. We further removed SNPs with a GoldenGate primer feasibility score ≤ 0.6, which also excluded failed primer designs (Supplementary Table 1). For the background variant set, we chose only one SNP per gene. For the fixed and outlier markers and candidate genes, we allowed one variant with the highest primer design score in each gene region (e.g. exon, intron, regulatory region defined as 5 kb surrounding a gene), which resulted in a higher density within the peak region on chromosome 18.

The exact location of each SNP in reference to genome version 2.5 (RefSeq Assembly ID on NCBI: GCF_000738735.1 as published in ⁸) and primer sequences are available in Supplementary Table 1. GoldenGate genotyping was performed on 520 samples at the SNP & SEQ Technology Platform, Uppsala. Cluster plots were automatically analysed using Illumina’s GenomeStudio software (v2011.1). We removed 31 individuals that had a call rate of 0%, and two additional duplicated individuals. We further removed 41 SNPs that had a missing call rate >25% (N = 37) or were monomorphic (N = 4). This resulted in a final set of 51 fixed, 735 outlier and 325 background genome-wide SNPs. For the fixed set, F_ST ranged from 0.766–1.000 (mean = 0.907) and 0.844–1.000 (mean = 0.970) between allopatric populations in the central and southern hybrid zone, respectively. For the outlier set, it ranged from -0.022–1.000 (mean = 0.199) and -0.050–1.000 (mean = 0.282) and for the background set from -0.022–0.378 (mean = 0.029) and -0.047–0.574 (mean = 0.100), respectively. One individual was genotyped three times to assess genotyping errors. No discordant genotype calls were observed among 2217 comparisons suggesting an accuracy of genotype calls above 99.95%. Of the allopatric populations, 40 individuals had been previously genotyped using the HaplotypeCaller in GATK (v3.3.0³²) after paired-end whole-genome sequencing on the HiSeq2000 (Illumina) platform³. Coverage ranged from 6.83× to 23.45× (average = 11.37×, median = 10.64×). Five of those individuals were additionally genotyped using GoldenGate. In those, we found 104 inconsistencies out of 5501 genotypes (1.89%). For our final data set we kept the genotype calls from GoldenGate wherever possible, such that we had N = 522 individuals genotyped at N = 1111 polymorphic loci (average call rate of 99.32 %). The distribution of the 1111 SNPs in the genome is illustrated in Supplementary Fig. 3. Genotypic data is available in Supplementary Table 8.

Analyses

Hybrid indices and hybrid classes

We estimated maximum likelihood hybrid indices separately for the central and southern hybrid zone including all nest mates. Allopatric German, Polish and Swedish samples were used as ancestral reference populations in the central zone, Italian and different allopatric German populations for the southern zone, respectively. Previous work has shown that most of the genetic variation separating carrion and hooded crows clusters within a ~2 Mb region on chromosome 18⁸. Therefore, we estimated hybrid indices separately for SNPs on chromosome 18, containing the most ancestry informative SNPs, and the rest of the autosomal SNPs with the R-package introgress (v1.2.3³³) in R (v3.4.3³⁴). In the following, we refer to these hybrid indices as chr18 and GW (genome wide), respectively. A hybrid index of 0 represents a purebred carrion crow and a hybrid index of 1 a purebred hooded crow.

Using the NewHybrids software (v2.0+ Developmental, July/August 2007) we further estimated the posterior probability of an individual belonging to one out of six distinct genealogical classes (purebred carrion or hooded crows, F1- or F2-hybrids, and backcrosses to carrion or hooded crows) by simulating the different genotype frequency classes³⁵. We used the parallelnewhybrid (v0.0.0.9002³⁶) R-package to call NewHybrids with uninformative Jeffreys-type priors for the estimation of allele frequencies (ϴ) and mixing proportions of genotype frequency classes (π). We ran NewHybrids separately for the central and southern hybrid zone and separately for SNPs on chromosome 18 and the rest of the autosomal SNPs, discarding the first 20,000 generations as burn-in and estimating parameters from the following 200,000 MCMC algorithm iterations. Assignment efficiency, accuracy and overall performance were assessed by simulating multigenerational hybrid data sets³⁷. For that we used the hybriddetective (v0.1.0.9000³⁸) R-package implementing the HybridLab algorithm³⁹. Using the genotype data of the allopatric samples (see above), we simulated four times N = 500 purebred (that is carrion or hooded crows), N = 250 F1, N = 125 F2 and N = 125 BC individuals. The relative frequencies of the different genealogical classes were matched to those inferred empirically by NewHybrids, the performance of which is known to be affected by the proportion of hybrids in the sample³⁷. Following the procedure used for the empirical data, we ran all simulations separately for the central and southern hybrid zone and separately for SNPs on chromosome 18 and the rest of the autosomal SNPs. We then combined these simulated data sets with the allopatric samples and ran NewHybrids using the same settings as for the empirical data. We merged the results of the four simulated data sets (4 × central autosomal SNPs, 4 × southern autosomal SNPs, 4 × central chromosome 18 SNPs, 4 × southern chromosome 18 SNPs) and estimated NewHybrids’ assignment efficiency, accuracy and overall performance. We used the simulated data to select the optimal posterior probability cut-off that maximized the accuracy of NewHybrids genealogical class assignment. The maximum was at an overall posterior probability ≥0.995 across all genealogical classes for both SNP sets in the two hybrid zones. Using this cut-off, NewHybrids assigned 99.14% of purebred, 85.98% of F1, 85.75% of F2 and 95.98% of backcrosses correctly.

For the empirical data, NewHybrids’ genealogical class assignments and introgress’ hybrid indices were in good correspondence using the SNPs on chromosome 18 (Supplementary Fig. 12). We visually inspected graphical representations of marker ancestry across chromosome 18 in all F1-hybrid individuals (using introgress’ mk.image() function), which showed that these individuals were indeed heterozygous at the outlier loci on chromosome 18. Using the SNPs on chromosome 18, NewHybrids assigned all individuals into the genealogical classes with a posterior probability of 1, and most of the individuals were either classified as purebreds (65.8% and 72.9% in the central and southern hybrid zone, respectively) or as F1-hybrids (30.3% and 25.8% in the central and southern hybrid zone, respectively, removing nest mates; Supplementary Table 9). We call these clear diplotypes based on chromosome 18 chr18_DD, chr18_LL and chr18_DL, respectively.

We additionally used these plots to identify inter-specific recombination breakpoints along chromosome 18 by taking change in haplotype ancestry as evidence for a recombination event (Supplementary Fig. 13). We only considered changes in haplotype tracts of at least three SNPs, single SNP changes were discarded as genotype errors or variation. Assuming randomness of cross-over locations, haplotype tracts of the exact same lengths shared between nest mates are likely to be identical by descent and were only counted as a single recombination event.

Phenotype characterization and genome-wide associations (GWA)

Digital photographs of the ventral and dorsal side were taken for all birds from Graz and for around half of the birds from the southern hybrid zone (N = 109 pictures of 18 individuals from Graz and N = 456 pictures of 111 individuals from the southern hybrid zone) for scoring the amount of grey and black in the feathering (see below). For each of the 129 individuals we scored seven plumage patches on the ventral side and four regions on the dorsal side for the amount of grey and black in the feathering (Fig. 1a, Supplementary Fig. 4). Each patch was scored as 0 = pure grey, 1 = dark grey or a mixture of grey and black feathers or 2 = pure black by the same person (U.K., Supplementary Table 10). Because some plumage patches could not be scored reliably, we imputed missing data (2.18%) using the missMDA R-package (v1.11⁴⁰). Individual measurements were then summarized using a principal component analysis (PCA) as implemented in the FactoMineR R-package (v1.39⁴¹). All eleven variables loaded positively on principal component 1 (PC1), which explained 78.22% of the variance in pigmentation pattern and intensity. PC2 explained 7.66% of the variance and separated the four belly quadrants and the mantle from the other plumage patches. In order to assess the objectivity of scoring, we estimated inter-observer repeatability of the colour PC1 and PC2 scores by measuring a subset of individuals by a second person (N.S., N = 105 individuals). We estimated the repeatability of PC1 and PC2 scores using the rptGaussian() function from the rptR R-package (v0.9.21⁴²). Both PC1 and PC2 scores were highly repeatable between observers (PC1: r ± SE = 0.965 ± 0.007, P = 4 × 10^-64; PC2: r ± SE = 0.646 ± 0.057, P = 1 × 10^-14 using N = 10.000 parametric bootstraps for interval estimation).

We performed two GWA studies, using either colour PC1 or colour PC2 as dependent variables and each SNP as a covariate (coded as 0, 1, or 2 copies of the minor allele) using one degree of freedom (additive effect). Models were fitted with the qtscore() function and a Gaussian error structure in the GenABEL R-package (v1.8-0⁴³). Since SNPs showed varying degrees of linkage disequilibrium, we used the simpleM-algorithm with default settings to estimate the effective number of tests performed⁴⁴ and used this estimate for controlling the genome-wide type I error rate.

We then fitted all SNPs showing a significant additive main effect on colour PC1 (3 SNPs on chromosome 1, 1 SNP on chromosome 1A, 166 SNPs on chromosome 18) as factors to test for additional dominant gene action and tested epistatic interactions between all of them (3 × 1 + 3 × 166 + 1 × 166 = 667 interactions). For significance testing, we performed likelihood ratio tests comparing models with and without the dominant gene action or interaction. Because both chromosome 18 and the three SNPs on chromosome 1 effectively behave as one locus, we reduced our model to include only the ancestry on chromosome 18 and the most significant SNP on chromosome 1 in our final model. We selected the best model based on Akaike’s information criterion using ΔAIC ≥ 2 as selection threshold⁴⁵.

Under Hardy-Weinberg-equilibrium the variance explained by a SNP is calculated as V_SNP = 2 × p × (1 - p) × β², where p is the minor allele frequency and β the estimated slope for the SNP-effect (i.e. the average effect of an allelic substitution⁴⁶) from the above regression models. To obtain the proportion of phenotypic variance explained by a SNP, V_SNP is divided by the phenotypic variance, which is equivalent to the multiple R-squared (r²) from an ordinary least squares regression model. In our study populations, however, most SNPs were not in Hardy-Weinberg-equilibrium and we thus used the multiple R-squared (r²) from the regression models to get an estimate of the variance explained by additive, dominance or interaction effects.

Geographic clines

We used one-dimensional geographic cline analysis to assess whether genetic variation in the outlier region of chromosome 18 and the significant loci from the GWAS showed a signal of contemporary divergent selection. Several sampling locations did not fully coincide with the transect line. In these cases we collapsed the two-dimensional geographic coordinates using principal component analyses for both the central and southern hybrid zone separately (without the allopatric populations) and reconstructed coordinates using PC1 only. We then calculated great-circle (orthodromic) distances in km between these reconstructed coordinates. This essentially corresponds to a perpendicular projection of sampling locations onto a transect minimizing the distance to the sampling locations. We added the allopatric populations by calculating their great-circle distances from the respective ends of the sampling transect. We then fitted geographic cline models using the hzar (v0.2-5⁴⁷) package in R using the mean chr18 and GW hybrid indices per sampling location as the dependent variable and geographic distance between sampling locations (in km) as the predictor. We included only one individual per nest in order to reduce pseudoreplication (i.e. we kept N = 116 individuals in the central and N = 273 individuals in the southern hybrid zone). In hzar, cline models as described in ^17,48 are fitted using the Metropolis-Hastings Markov chain Monte Carlo (MCMC) algorithm^49,50. Model selection between the null model and three cline models (I–III) varying in the number parameters was based on Akaike’s information criterion using ΔAIC ≥ 2 as selection threshold⁴⁵. In model I, we fixed the minimum and maximum hybrid index frequencies p_min and p_max to their minimum and maximum observed mean values (scaling = fixed) and fitted a model without exponential tails (tails = none). Thus, in model I only the sigmoid cline function is fitted and the two parameters cline centre and cline width are estimated. In model II we additionally estimated the minimum and maximum hybrid index frequencies p_min and p_max (scaling = free) but did not include exponential tails (tails = none). In model III we also estimated the minimum and maximum hybrid index frequencies p_min and p_max and fitted exponential tails with independent size (δ_left, δ_right) and shape (τ_left, τ_right) parameters (tails = both). In one case (southern hybrid zone, mean chr18 hybrid index) ΔAIC was 1.89 and we present results of model I which had the non-significant better fit.

Genomic clines

Genomic cline models use the hybrid index instead of the geographic distance to detect loci that may be subject to selection. We fitted genomic cline models both with the gghybrid R-package (v0.0.0.9000⁵¹) and with bgc (v1.03⁵²) for both the central and the southern hybrid zone. gghybrid allows fitting hybrid indices that have been estimated using only a subset of the markers. Previous work suggests that markers in the outlier region on chromosome 18 are in significant linkage disequilibrium and putatively under positive selection⁸. To test against a genome-wide background of neutral genetic variation we fitted models in gghybrid with hybrid indices estimated from all SNPs excluding chromosome 18 (but including chromosome Z). gghybrid uses the same method for estimating hybrid indices as the introgress R-package (with an additional prior⁵³) and hybrid indices were highly correlated between the two packages (r ≥ 0.99). We then fitted genomic cline models including only a single individual per nest and the allopatric populations, discarding the first 10,000 iterations as burn-in and estimating parameters and posterior probabilities (P-values) from the following 40,000 MCMC iterations. gghybrid fits Fitzpatrick's logit-logistic cline function ⁵⁴ and estimates parameters v and u. Parameter v is always positive and higher values indicate steeper clines. Parameter u is related to the centre of the cline, i.e. the hybrid index at which the allele frequencies are half way between those of the allopatric parental populations.

The Bayesian genomic cline model implemented in bgc derives hybrid indices internally using all markers and then fits the Barton cline model¹⁷. Thus, the bgc hybrid indices integrate over both the hybrid indices estimated from SNPs on chromosome 18 and those estimated using all remaining autosomal SNPs. bgc derives parameters and no P-values and parameter estimates should be unaffected by related individuals. Thus, we included all nest mates in these models. bgc models were fitted with the ICARrho model for linked loci, discarding the first 250,000 generations as burn-in, estimating parameters from the following 500,000 MCMC algorithm iterations and recoding every 100th value and default parameters otherwise. We ran two independent MCMC simulations per hybrid zone and combined the output of the two chains. We estimated means and 95% confidence intervals for the slope parameter β, which describes the rate at which the probability of ancestry transitions. Positive values indicate steeper clines and a reduced rate of introgression⁵⁵. Since we used all SNPs including those under divergent selection on chromosome 18, β is downward biased due to the fact that SNPs are densely spaced on chromosome 18, show the strongest clinal variation and contribute disproportionately to the hybrid index (correlation of internally derived hybrid index with chr18: r = 0.94 and GW: r = 0.84 for the central hybrid zone and chr18: r = 0.90 and GW: r = 0.82 for the southern hybrid zone). For visual representation we thus normalized the distribution of β-estimates by the median. Outliers were classified as loci falling outside the empirical 95% quantile distribution. The analysis was repeated excluding SNPs on chromosome 18, in which case the correlation of internally derived hybrid indices with GW was much higher (r = 0.97 for the central hybrid zone and r = 0.96 for the southern hybrid zone).

Population genetic analyses

Ancestral state reconstruction

We determined the ancestral state of all SNPs in the GoldenGate assay using the data and same methods as in ⁸ for five genomes of the American crow (Corvus brachyrhynchos), two rook genomes (Corvus frugilegus) and two genomes of the jackdaw (Corvus monedula). 80.72% of the sites were fixed in all three outgroup species (73.29% in ⁸) and thus informative for the ancestral state in hooded and carrion crows. 11.84% of the sites were polymorphic in American crows but fixed for the same allele in rooks and jackdaws, which most likely represent ancestral polymorphisms prior to the divergence of American crows from hooded and carrion crows. These sites are thus uninformative in respect to the ancestral state in hooded and carrion crows. 3.23% of all sites were fixed for the same allele in American crows and rooks but polymorphic in jackdaws. This could be due to either a mutation in the lineage leading to jackdaws (homoplasy) or incorrectly mapped paralogues, and these sites were thus excluded. We also removed 1.88% of the sites that were fixed for alternative alleles in two species and polymorphic in the third or biallelic in two or more species. Finally, we discarded sites that were polymorphic in rooks and fixed for the same allele in jackdaws and American crows (1.43%), those that were triallelic (0.45%) or those where genotype information was missing (0.45%). For a summary of the data see Supplementary Table 11.

F-statistics estimation

F_ST values were estimated from genotypes derived from whole-genome sequencing data³ using PLINK (v1.90b4.4⁵⁶). We used the allopatric samples of carrion and hooded crows from the central and southern hybrid zone independently as input populations. F_IS values were estimated from chromosome 18 genotypes of individuals sampled in the central and southern hybrid zone and that were assigned either a hooded or carrion crow or F1-hybrid ancestry. For that we used the R-package hierfstat (v0.04-22⁵⁷).

Principal component analyses (PCA)

To better understand the partitioning of genetic variation along chromosome 18 (see ^58,59), we performed PCA using all individuals from the two hybrid zones (N = 173 individuals and N = 236 individuals for the central and southern zone, respectively) and all SNPs on chromosome 18 for which we could resolve the ancestral state and that had a missing call rate smaller 0.05 and a minor allele frequency larger than 0.05 (N = 204 SNPs and N = 193 SNPs for the central and southern zone, respectively). Discrete clusters of genetic variation provide evidence for divergent haplotypes that are non-recombining (for example due to an inversion). In case of two major haplotypes, we expected principal component 1 (PC1) to classify individuals by diplotype: homozygous individuals are expected to cluster at both ends of the PC1 distribution with heterozygous individuals in between. We further expect the squared principal component loadings of PC1 to reflect F_ST (termed communality h²; ⁶⁰), if the haplotype structure on chromosome 18 was the same between the allopatric populations (used for F_ST) and the individuals from the hybrid zone. By using the ancestry information, the principal component loadings can be polarized. This provides a directional measure of population differentiation, which should provide information on the direction of positive selection (increase in the proportion of derived variants in the target population). PCA was performed using the R package SNPRelate (v1.12.1⁶¹).

One sentence summary.

Admixture mapping identifies epistatic interaction of major-effect pigmentation loci as the molecular genetic basis of prezygotic isolation.