Analysing population genomic data from killer whale ecotypes, which we estimate have globally radiated within less than 250,000 years, we show that genetic structuring including the segregation of potentially functional alleles is associated with socially inherited ecological niche. Reconstruction of ancestral demographic history revealed bottlenecks during founder events, likely promoting ecological divergence and genetic drift resulting in a wide range of genome-wide differentiation between pairs of allopatric and sympatric ecotypes. Functional enrichment analyses provided evidence for regional genomic divergence associated with habitat, dietary preferences and post-zygotic reproductive isolation. Our findings are consistent with expansion of small founder groups into novel niches by an initial plastic behavioural response, perpetuated by social learning imposing an altered natural selection regime. The study constitutes an important step towards an understanding of the complex interaction between demographic history, culture, ecological adaptation and evolution at the genomic level.
The interplay between ecology, culture and evolution at the level of the genome remains poorly understood1. The ability to adapt to novel ecological conditions through behavioural plasticity is thought to be able to buffer natural selection pressures and promote rapid colonization of novel niches1. However, by perpetuating exposure to a novel environment, stable cultural transmission of behaviour can also provide an opportunity for natural selection to act on adaptive genomic variation. Examples of genomic adaptation in humans during the period of recent ecological and cultural diversification and consequent demographic expansion are well illustrated1,2,3. For example, the Inuit of Greenland descend from a small founder population that split from an East Asian source population and successfully colonized the extreme climatic conditions of the Arctic environment through culturally transmitted methods of hunting marine mammals and genetic adaptation to a cold climate and hypoglycaemic lipid-rich diet4. However, our understanding of the complex interaction between ecology, culture, adaptation and reproductive isolation at a genome-wide level has long suffered from deficiency of genome-wide data, and, conceptually, from the almost-exclusive focus on these processes in humans and thus a lack of comparative data from other species5.
Killer whales (Orcinus orca) are the largest species in the dolphin family (Delphinidae) and, together with humans, are one of the most cosmopolitan mammals, being found in all ocean basins and distributed from the Antarctic to the Arctic6. This top marine predator consumes a diverse range of prey species, including birds, fish, mammals and reptiles6. However, in several locations killer whales have evolved into specialized ecotypes, with hunting strategies adapted to exploit narrow ecological niches6,7,8,9,10 (see Supplementary Notes for more detailed information on the natural history of killer whale ecotypes). For example, in the North Pacific, two sympatric ecotypes coexist in coastal waters: the mammal-eating (so-called ‘transient’) ecotype and fish-eating (so-called ‘resident’) ecotype7,8. This ecotypic variation is stable among multiple subpopulations of the transient and resident ecotypes across the North Pacific that diverged from common ancestral matrilines ∼68 and 35 KYA, respectively11. A highly stable matrilineal group structure and a long post-menopausal lifespan in killer whales is thought to facilitate the transfer of ecological and social knowledge from matriarchs to their kin12, and thereby perpetuate the stability of ecotypic variation in killer whales13. In the absence of a definitional consensus and for the purposes of investigating how cultural phenomena interact with genes, culture has been broadly defined as ‘information that is capable of affecting individuals’ behaviour, which they acquire from other individuals through teaching, imitation and other forms of social learning’1. Several studies have argued that behavioural differences among killer whale ecotypes are examples of culture in this broader sense of the term13,14. However, this behavioural variation among ecotypes likely results from ecological, genetic and cultural variation and the interaction between them, rather than a single process explaining all behavioural variance15. Killer whales, therefore, offer a prime example of how behavioural innovation perpetuated by cultural transmission may have enabled access to novel ecological conditions with altered selection regimes, and thus provide an excellent study system for understanding the interaction between ecological and behavioural variation, and genome-level evolution.
Results and Discussion
We generated whole-genome re-sequencing data of 48 individuals at low coverage and accessed high-coverage sequencing data from two more individuals16,17 to investigate patterns of genomic variation among killer whale ecotypes. The samples represent five distinct ecotypes that, based on phylogenetic analysis of mitochondrial genomes11, include some of the oldest and youngest divergences within the species (Fig. 1). The dataset incorporated 10 individuals each of the transient and resident ecotypes that occur in sympatry in the North Pacific; and from Antarctic waters, 7 ndividuals of a large mammal-eating form (type B1), 11 individuals of a partially sympatric, smaller form which feeds on penguins (type B2), and 10 individuals of the smallest form of killer whale, which feeds on fish (type C) (Fig. 1). A total of 2,577 million reads uniquely mapped to the 2.4-Gbp killer whale reference genome16 (Supplementary Fig. 1) for which a chromosomal assembly was generated for this study, so that approximately 50% of the autosomal regions of each individual were sequenced at ≥2 × coverage (Supplementary Table 1). Subsequent data analyses used methods that account for uncertainty in the assignments of genotypes, enabling accurate inferences to be drawn from low-pass next-generation sequencing data18. Comparisons of estimated population genomic metrics such as genome-wide and per-site FST indicated that estimates from our low-coverage data were highly consistent with published high-coverage restriction site associated DNA sequencing (RAD-seq)19 and single-nucleotide polymorphism (SNP)-typing11 data (Supplementary Fig. 2 and Supplementary Table 2) and thus confirmed the robustness of our estimates.
Time to most recent common ancestor
We estimated a time to most recent common ancestor (TMRCA) of ∼126–227 KYA from the accumulation of derived mutations at third-codon positions for the most divergent killer whale lineages compared here (Supplementary Table 3) and based on the 95% highest posterior density interval of the mutation rate estimate20. This equates to ∼4,900–8,800 generations and indicates a rapid diversification over a timescale comparable to the diversification of modern humans21. We caution that for these age estimates we rely on mutation rate estimates derived from interspecific comparisons among odontocetes, and that, therefore, these estimates of TMRCA are at best approximate. Further, the demographic history and any gene flow between ecotypes will have an influence on the sharing of derived mutations and hence this estimate. However, we do expect that the estimate will be within the correct order of magnitude. Our estimated TMCRA overlaps with a recent RAD marker study22, which estimated a TMRCA of 189 KYA (only a point estimate was reported by these authors) scaling by a mutation rate 1.21 times higher than we have used here. The estimated TMRCA of a global data set of killer whales based on non-recombining mitochondrial genomes has been estimated at 220–530 KYA (ref. 11), older than our estimate based on nuclear genomes. Male-mediated gene flow has therefore continued after matrilineal lineages have diverged.
Genetic differentiation and divergence
Despite this recent shared ancestry, substantial genome-wide differentiation and divergence had accrued between all pairs of ecotypes included in this study (Fig. 2 and Supplementary Table 2). At K=5 populations, a maximum-likelihood-based clustering algorithm23 unambiguously assigned all individuals to populations corresponding to ecotype (Fig. 2c), indicating that all ecotypes have been assortatively mating long enough to allow allele frequencies to drift apart. Pairwise genetic distances between individuals visualized as a tree indicate that segregating alleles are largely shared within an ecotype (Supplementary Fig. 3). Similarly, pairwise relatedness due to identity-by-descent, that is, genetic identity because of a recent common ancestor, was high within each ecotype. While the three Antarctic ecotypes still showed signs of recent relatedness, no shared recent identity-by-descent ancestry was detected between Antarctic and Pacific types or between the sympatric resident and transient ecotypes (Supplementary Fig. 4). The greatest differentiation (Supplementary Table 2) as visualized in the maximum likelihood graph (Fig. 2a) and PCA plot (Fig. 2b) was between the allopatric Pacific and the Antarctic ecotypes, while differentiation among Antarctic ecotypes was much lower than between the two Pacific ecotypes. Thus, our sampled populations allowed us to investigate the accrual of genomic differentiation along points of a continuum, acting as a proxy of sampling at different stages of the speciation process. The accrual of genome-wide differentiation (FST=0.09) between even the most recently diverged and partially sympatric ecotypes (Antarctic types B1 and B2) indicates that reproductive isolation quickly becomes established after the formation of new ecotypes. Thus, whole-genome resolution confirms that, even in sympatry, contemporary gene flow occurs almost exclusively among individuals of the same ecotype, allowing genomic differentiation to build up between ecotypes so that within an ocean basin ecological variation better predicted genetic structuring than geography.
To better understand and visualize the complexity of the ancestry of killer whale ecotypes, we reconstructed the genetic relationships among ecotypes in the form of a maximum likelihood graph (Fig. 2a), representing the degree of genetic drift and modelling both population splits and gene flow using the unified statistical framework implemented in TreeMix (ref. 24). The inferred migration edges were supported by the three-population (f3) and D-statistic (ABBA-BABA)25 tests, which can provide clear evidence of admixture, even if the gene flow events occurred hundreds of generations ago26. These population genomic methods test for asymmetry in the covariance of allele frequencies that indicate that the relationships among populations are not fully described by a simple bifurcating tree model.
The three approaches were consistent in inferring migration from source populations that share ancestry with the North Pacific resident and North Atlantic ecotype into the transient ecotype (Fig. 2a, Supplementary Figs 5 and 6, Supplementary Tables 4 and 5 and Supplementary Data 1). The genomes of transients are therefore partly derived from at least one population related to the Atlantic and resident ecotypes, (but not necessarily these populations, that is, the source could be an unsampled ‘ghost’ population). The asymmetrical two-dimensional (2D)-site frequency spectrum (SFS) also implies directional gene flow from a population ancestrally related to the residents into the transient ecotype27 (Fig. 2e). Given the extent of the inferred demographic bottlenecks during founder events (see section below), and the expected consequential shift in the SFS28, it seems likely that this admixture would have occurred during or after the founder bottleneck in the transient ecotype, otherwise it would be expected to be less correlated with the resident SFS. The sequencing of more populations is expected to shed further light on this episode of ancient admixture.
TreeMix, the three-population and D-statistic tests inferred that type B1 is admixed and derives from at least two populations related to both types B2 and C (Fig. 2a, Supplementary Tables 4 and 5 and Supplementary Data 1). Conducting D-statistic tests on proposed tree-like histories comprising combinations of 16 genome sequences that included the North Atlantic sequence, we found that Antarctic types B1 and B2 shared an excess of alleles with the three Northern Hemisphere ecotypes (Supplementary Figs 7 and 8). This shared ancestry component between Northern and Southern hemisphere ecotypes was not detected in type C, suggesting a relatively recent admixture event after type C split from the shared ancestor of types B1 and B2. Other signals of ancient admixture among populations were also detected and are reported in the tables in the Supplementary Material.
As our results suggest that killer whale ecotypes have diversified rapidly from a recent ancestor, and given the importance of the relationship between effective population size (Ne) and the rate of evolution29, we conducted analyses to reconstruct their demographic history. Applying the pairwise sequential Markovian coalescent (PSMC) approach21 to two high-coverage (≥20 × ) autosomal assemblies, a North Atlantic female and a North Pacific resident male, refining the methodological approach of a previous analysis of these genomes (Supplementary Figs 9 and 10), we recovered a similar demographic trajectory to that previously reported17 (Fig. 3a). The inference of the timing of these demographic declines is dependent on the assumed mutation rate (μ); however, across a range of sensible estimates of μ, the declines broadly fall within the Late Pleistocene17. This was previously interpreted as evidence for demographically independent population declines in each ocean, driven by environmental change during the Weichselian glacial period17. However, this inference assumes that each PSMC plot tracks the demographic history of a single unstructured panmictic population30. The y axis of the PSMC plot is an estimation of Ne derived from the rate of coalescence between the two chromosomes of a diploid genome. However, in the presence of population structuring, regions of the two chromosomes will coalesce less frequently as their ancestry may derive from different demes or subpopulations; the rate of coalescence can thus be similar to a single population with large Ne. PSMC estimates of Ne during population splits can therefore be greater than the sum of the effective sizes of the subpopulations, dependent on the number of subpopulations and the degree of connectivity (that is, cross-coalescence) between them21,30. In fact, the results presented here and previously published17,19,22 indicate that throughout the Weichselian glacial period there were multiple population splits, both between and within ecotypes, including the splitting of the two lineages included in the PSMC analyses just at the point of the change in inferred Ne (Fig. 3a and Supplementary Fig 11). Therefore, the PSMC plots will be strongly influenced by these changes in structure, and the changes in inferred Ne are not necessarily associated with population declines (see Supplementary Fig. 12).
To investigate the influence of connectivity on our PSMC plots further, we inferred ancestral Ne of the diploid X-chromosome (NeX) of the Atlantic female and directly compared with the autosomal fraction (NeA) of the genome presented in Fig. 3a. From ∼300,000 to 130,000 years BP during the Saalian glacial period, the inferred NeX ranged from 0.58 to 0.79 of the NeA (Fig. 3b), which lies within expectation for demographically stable mammalian species31. During the first part of the Weichselian glacial period, NeX markedly declined, reaching a minimum at ∼30,000–50,000 years BP. The timing of this bottleneck in NeX overlaps with the stem age of the mitochondrial clade for this Atlantic population11, that is, consistent with almost all mitochondrial diversity being lost in this lineage during this period. Conversely, this is concurrent with the peak estimate of autosomal NeA inferred by PSMC, and ratio of NeX/NeA falls to ∼0.3 at this NeX minimum and then recovers within 1,000 generations to >0.75 (Fig. 3b). Simulated demographic bottlenecks of several hundredfold reduction result in a disproportionate loss of NeX, attributed to the difference in the inheritance mode of each marker, and the ratio of NeX/NeA can reach less than 0.3 (ref. 31). Following the bottleneck, NeX recovers more rapidly than NeA and the ratio of NeX/NeA can exceed 0.75 during this recovery phase31. The concordance of the timing of the bottleneck in the X-chromosome and the stem age of the mitochondrial genome suggest an underlying demographic process, rather than a strong selection on the X-chromosome or some other factors driving the mutation rate31. A sex-biased process such as primarily male-mediated gene flow between demes could further influence the ratio of NeX/NeA (ref. 32).
To estimate the demographic histories from our population genomic data, we produced ‘stairway’ plots using composite likelihood estimations of theta (θ) for different SNP frequency spectra associated with different epochs, which are then scaled by the mutation rate to estimate Ne for each epoch33. Using this method we reconstructed a demographic history from our population genomic data for the resident ecotype that was comparable to the PSMC plot from a single resident high-coverage genome, both methods identifying a decline in Ne starting at ∼60 KYA (Fig. 3a,c). The stairway plots for the transient ecotype and type C showed the same pattern as the resident of a decline to a bottlenecked population with an Ne of <1,000 and a subsequent expansion (Fig. 3c). The bottlenecks did not occur simultaneously, as might be expected in response to a global environmental stressor during a glacial cycle, but instead were sequential (Fig. 3c). In each case, the timing of the demographic bottleneck overlapped with the previously estimated timing of the stem age of the mitochondrial genome clades containing each ecotype11. Thus, within the transient, resident and combined Antarctic ecotypes, both mitochondrial and nuclear lineages coalesce back to these respective bottleneck events, consistent with genetic isolation of small matrifocal founder groups from an ancestral source population, followed by the subsequent expansion and substructuring of a newly established ecotype.
Overall, the population genetic analyses of the whole-genome sequences above shed light on the ancestry of killer whale ecotypes in unprecedented detail, highlighting a complex tapestry of periods of isolation interspersed with episodic admixture events and strong demographic bottlenecks associated with the founder events that gave rise to the resident, transient and ancestral Antarctic ecotypes.
Genome-wide landscape of genetic diversity and differentiation
Demographic bottlenecks during population splits and founding events, followed by subsequent demographic and geographic expansion, can produce rapid shifts in allele frequencies between populations34. The high levels of genome-wide differentiation (FST) between killer whale ecotypes across all genomic regions (Fig. 4a,b) are consistent with strong genetic drift following demographic expansion from small founding groups. Considering the low efficiency of selection in populations as small as the estimates presented here35, in which founder populations have an estimated Ne ranging from a few tens to hundreds, a genome-wide contribution of ecologically mediated divergent selection is neither necessary nor particularly likely to explain the observed shifts in allele frequencies in such a large number of loci. Consistent with this prediction, we find that differentiation is highest along the branches inferred by TreeMix to have experienced the most substantial genetic drift (Fig. 2a), that is, the branch to the ancestor of the Antarctic types and the branch to the resident ecotype (Supplementary Fig. 13). We therefore expect that only those beneficial alleles that have a strong favourable effect (that is, strength of selection (s)>1/2Ne) would have an increased fixation probability because of selection within these founder populations.
Much of the heterogeneity in the differentiation landscape was shared among pairwise comparisons (Fig. 4c). Since diversity (π) was not associated with mutation rate μ, as inferred from neutral substitution rate dS, (r=−0.1 to −0.24), the observed covariation in differentiation (FST), diversity (π) and absolute divergence (Dxy) between population pairs (Fig. 4c and Supplementary Fig. 14) is best explained by the shared local reduction of diversity by linked selection in the ancestral population36,37. This leads us to conclude that overall the landscape of genome-wide differentiation is a result of global genetic drift, regionally elevated by ancestral linked selection (for example, background selection or repeated selective sweeps shared among populations) independent of the evolutionary dynamics of the recently derived present-day ecotypes.
Genomic signatures of climate and diet adaptation
However, against this background of shared differentiation, there was evidence for genic divergence of putative functional relevance. The first targets of selection following ecotype diversification and the exploitation of new ecological niches are expected to be those that facilitate ecological specialization36. Once a certain level of reproductive isolation is reached, differences can accumulate in other (fast-evolving) genes involved in reproductive isolation effectively reducing hybrid mating38. Following this rationale, both individual gene associations and gene ontology (GO) enrichment analyses yielded several biological processes and candidate genes with putative functional roles in ecological specialization, local adaptation and reproductive isolation (Supplementary Tables 6–8). For example, in comparisons between ecotypes inhabiting the extreme cold of the Antarctic pack ice with ecotypes from the more temperate North Pacific (Fig. 1), we found the most significant enrichment in genes involved in adipose tissue development (GO:0060612, Fisher’s exact test: P=0.0015). Genes associated with adipose tissue development have previously been found to be evolving under positive selection in the polar bear39, suggesting a role for this process in rapid adaptation to a cold climate and lipid-rich diet.
Using the population branch statistic (PBS), which has strong power to detect recent natural selection40 and has allowed us to investigate allele changes along specific branches, we identified another candidate example where cold adaptation may play a role. The FAM83H gene showed a signature of selection (top 99.9% PBS values) and was found to contain four fixed non-synonymous substitutions derived in the Antarctic lineages based on the inferred ancestral state, which resulted in physicochemical changes including a hydrophobic side chain being replaced by a positively charged side chain. The keratin-associated protein encoded by the FAM83H gene is thought to be important for skin development and regulation through regulation of the filamentous state of keratin within cytoskeletal networks in epithelial cells, determining processes such as cell migration and polarization41. Skin regeneration is thought to be constrained in killer whales while inhabiting the cold waters around Antarctica because of the high cost of heat loss, and is thought to underlie rapid round-trip movements to warmer subtropical waters by Antarctic ecotypes42. The balance between skin regeneration and thermal regulation in Antarctic waters could be a major selective force requiring both behavioural42 and genomic adaptation (Supplementary Fig. 15).
Genes encoding proteins associated with dietary variation also showed a signature of selection (top 99.9% PBS values). For example, the carboxylesterase 2 (CES2) gene encodes the major intestinal enzyme and has a role in fatty acyl and cholesterol ester metabolism in humans and other mammals43. Two exons of the CES2 gene had among the top 99.9 percentile PBS values because of changes in allele frequencies (including two fixed non-synonymous amino-acid changes) along the branch to the Antarctic types (Supplementary Fig. 16). Similarly, genes in the top 99.9 percentile PBS values were enriched for carboxylic ester hydrolase activity (GO:0052689, Fisher’s exact test: P<0.0001) in the fish-eating resident ecotype. Biological processes enriched in the resident killer whale included digestive tract morphogenesis (GO:0048546, Fisher’s exact test: P=0.0022), and gastrulation with mouth forming second (GO:0001702, Fisher’s exact test: P=0.0024): associated with the formation of the three primary germ layers of the digestive system during embryonic development. These results overlapped with enriched GO terms identified by a previously published RAD-seq study, despite the relatively sparse sampling of 3,281 SNPs in that study19. Enrichment of these GO terms was largely driven by differentiation in a single exon in the GATA4 gene, which included a fixed non-synonymous substitution, sequenced in both studies.
Signatures of selection along branches leading to the two predominantly mammal-eating ecotypes included in this study, the North Pacific transient and Antarctic type B1, were found in genes that play a key role in the methionine cycle (Fig. 5). Methionine is an essential amino acid that has to be obtained through dietary intake, and is converted through trans-sulfurcation to cysteine via intermediate steps of catalysis to homocysteine44. Any excess homocysteine is re-methylated to methionine44. Diets with different protein contents, such as between killer whale ecotypes, will differ in their content of methionine, and the enzymatic cofactors involved in the metabolism of methionine and homocysteine, which include folate, vitamins B6 and B12 (ref. 44; hence why vegetarians often take vitamin B12 supplements). While different genes and different biological processes showed a signature of selection in each of these two mammal-eating ecotypes (Fig. 5), in both cases the candidate genes and processes were associated with the regulation of methionine metabolism, which results in the generation of cysteine. Successful hunting of mammal prey by killer whales would provide a sudden and rich source of dietary methionine. This fluctuating intake of protein may place more of a selective pressure on the regulation of the metabolism of methionine than does the consumption of fish by piscivorous ecotypes.
Rapid evolution of reproductive proteins
High (top 99.9 percentile) PBS values, largely driven by fixed non-synonymous amino-acid substitutions, were further estimated for several genes that encode proteins associated with reproductive function, including testis development, regulation of spermatogenesis, spermatocyte development and survival, and initiating the acrosome reaction of the sperm (for example, PKDREJ, RXFP2, C9orf24, SPEF1, TSSK4, DHH and MMEL1). Reproductive proteins such as PKDREJ, in which we found two fixed non-synonymous substitutions derived in the ‘resident’ ecotype, are known to diverge rapidly across taxa, and because of their functional role in fertilization are emerging as candidates for the post-zygotic component of the speciation process38.
Overall, our results indicate that the processes underlying genomic divergence among killer whale ecotypes reflect those described in humans in several respects. Behavioural adaptation has facilitated the colonization of novel habitats and ecological niches. Founder effects and rapid formation of reproductive isolation, followed by population expansion, have promoted genome-wide shifts in the frequency of alternative alleles in different ecotypes due to genetic drift. Demographic changes during founder events and subsequent expansions can also influence cultural diversity45,46, and may have had a role in reducing within-ecotype cultural diversity and promoting cultural differentiation between ecotypes. As with studies on modern humans, it is difficult to demonstrate a causal association between cultural differences and selection on specific genes1; however, our findings of divergence in genes with putative functional association with diet, climate and reproductive isolation broadly imply an interaction between genetically and culturally heritable evolutionary changes in killer whale ecotypes. Given these findings, the almost-exclusive focus on humans by studies of the interaction of culture and genes5 should be expanded, and exploration of culture–genome coevolution models in suitable non-human animal systems encouraged.
Skin biopsies from free-ranging killer whales were collected using projected biopsy darts, concurrent with the collection of photographic, photogrammetry and behavioural data, allowing sampled individuals to be binned to ecotype/morphotype a priori to genetic analyses. Most samples were selected from separate collection dates and identified groups (when known) to minimize chances of collecting close relatives or replicate individuals. The sample set included 10 ‘resident’, 10 ‘transient’, 7 type B1, 11 type B2 and 10 type C killer whales.
Genomic DNA library building and sequencing
DNA was extracted using a variety of common extraction methods as per ref. 11. Genomic DNA was then sheared to an average size of ∼150–200 bp using a Diagenode Bioruptor NGS. Illumina sequencing libraries were built on the sheared DNA extracts using NEBNext (Ipswich, MA, USA) DNA Sample Prep Master Mix Set 1 following Meyer and Kircher47. Libraries were subsequently index-amplified for 15 cycles using Phusion High-Fidelity Master Mix (Finnzymes) in 50-μl reactions following the manufacturer’s guidelines. The libraries were then purified using the MinElute PCR purification kit (Qiagen, Hilden, Germany). The DNA concentration of the libraries was measured using a 2100 Bioanalyzer (Agilent Technologies, CA, USA); these were then equimolarly pooled by ecotype and each ecotype pool was sequenced across five lanes of an Illumina HiSeq 2000 platform using single-read 100-bp chemistry (that is, a total of 25 lanes).
Read trimming and mapping
A high-quality, 2,249-Mb reference killer whale genome assembly (Oorca1.1, GenBank: ANOL00000000.2, contig N50 size of 70.3 kb, scaffold N50 size of 12.7 Mb)16 was used as a mapping reference. For the purpose of this study, the genome was masked for repetitive elements using RepeatMasker48 and the Cetartiodactyl repeat library from Repbase49. Repetitive elements constitute 41.32% of the killer whale reference genome (929,443,262 sites). A further 80,599 sites were identified as mitochondrial DNA transposed to the nuclear genome (numts) and were masked accordingly. The final assembly was then indexed using BWA v. 0.5.9 (ref. 50) to serve as the reference for read-mapping.
Illumina HiSeq 2000 reads from each individual were processed with AdapterRemoval51 to trim residual adapter sequence contamination and to remove adapter dimer sequences as well as low-quality stretches at 3′ ends. Filtered reads >30 bp were then mapped using Burrows-wheeler aligner (BWA), requiring a mapping quality greater than 30. Clonal reads were collapsed using the rmdup function of the SAMtools (v. 0.1.18) suite52. Ambiguously mapped reads were also filtered out using SAMtools. Consensus sequences were then reconstructed in Binary sequence/Alignment Map file format. To ensure that all repeat regions, which have the potential to bias population genetic inference, were removed, the per-site coverage was calculated across all 48 individuals. Inspection of the data suggested that sites with a total depth (including data from all 48 individuals) of >200 × reads were likely to be unmasked repeats. We therefore further masked these regions, which constituted an additional 0.14% of the genome (3,241,923 sites), resulting in a total masking of 932,685,185 sites (41.46% of the genome).
Ancestral state reconstruction
The ancestral state for each site was inferred by mapping whole-genomic Illumina sequencing reads of the bottlenose dolphin (Tursiops truncatus, Short Read Archive accession code SRX200685)16 against the killer whale reference genome using BWA read mapper as above. The consensus sequence was called using SAMtools, and ambiguous bases were masked with N's. The ancestral state could be inferred for 2,206,055,540 (98.1%) of 2,249,565,739 bases.
Inferring time to most recent common ancestor
Nine of the highest coverage killer whales were selected from our data set, which included two individuals each of the resident, transient, type B2 and type C ecotypes, and a type B1 individual. Estimates of the TMRCA were based on the number of derived transitions and number of derived transversions at third-codon positions. Exclusively considering mutations at third-codon sites was expected to minimize the impact of incomplete purifying selection, which can lead to overestimation of the substitution rate on short timescales. However, some mutations at third codons are non-synonymous, notably more so for transversions than for transitions, and putatively ephemeral transversions may therefore result in the overestimation of TMRCA. The lower rate of transversions, compared with transitions, is expected to minimize the impact of recurrent mutations at the same site, which could result in an underestimation of TMRCA based on transitions. Therefore, the expectation is that our estimate of TMRCA based on transversions may be upwardly biased, and our estimate based on transitions may be downwardly biased.
From a total of 4,781,830 third-codon positions (reduced to 3,127,876 when sites with missing data in any of the nine individuals were masked), 7,547 were inferred to be transversions and 12,784 were inferred to be transitions (with a minor allele frequency (MAF) cutoff of 0.1 so as to exclude potential sequencing errors) from the ancestral state (inferred from comparison with the dolphin genome). Of these, 7,120 transversions and 11,176 transitions were fixed in the killer whale, and therefore were inferred to have occurred along the branch from the killer whale/bottlenose dolphin ancestor and the MRCA of the killer whales (or because of incorrect inference of the ancestral state); moreover, 421 transversions and 1,608 transitions occurred within one or more of the killer whale lineages. A further six sites were inferred to have undergone a transversion from the ancestral state in at least one of the killer whale lineages, but had derived transitions in at least one other killer whale lineage. The Ts/Tv ratio of derived mutations at third-codon positions was therefore estimated to be 3.8 within the killer whale clade.
The proportion of derived mutations at third-codon positions found in one individual and shared in another individual is expected to decrease in comparisons between individuals from populations that diverged longer ago, as the probability that the mutation occurred within just one population following the split increases. The proportion of derived transversions and transitions at third-codon positions inferred within the type B1 individual that were shared with each of the other eight individuals was measured. The two resident individuals shared the least number of derived transversions with the type B1 individual (Supplementary Table 3). The results were highly consistent between individuals of the same ecotype (Supplementary Table 3). The mean rate of nucleotide evolution estimated for odontocetes of 9.10 × 10−10 substitutions per site per year (95% highest posterior density interval: 6.68 × 10−10, 1.18 × 10−9)20 was then scaled by our estimate of the Ti/Tv ratio of 3.8 at third-codon positions within our killer whale data set and was used to predict the time taken to accumulate 124 derived transversions and 465 transitions at third-codon positions that we inferred had been derived in type B1 since splitting from a shared ancestor with the resident ecotype.
An individual-based assignment test was performed to determine whether the ecotypes to which each individual had been assigned a priori, based on observed behaviour and/or morphological characteristics at the time of sampling, represented discrete gene pools in Hardy–Weinberg equilibrium. Since the 48 genomes generated for this study had an average sequencing depth of 2 × , genotypes can only be called with very high uncertainty. Therefore, NGSadmix23, a maximum likelihood method that bases its inference on genotype likelihoods (GLs) and in doing so takes into account the uncertainty in the called genotypes that is inherently present in low-depth sequencing data, was employed. The method has been demonstrated, using simulations and publicly available sequencing data, to have great accuracy even for very low-depth data of less than twofold mean depth23. GLs were estimated using the SAMtools method52 implemented in the software package ANGSD53. NGSadmix was run with the number of ancestral populations K set to 3–5. For each of these K values, NGSadmix was re-run multiple times with different seeds in order to ensure convergence. Sites were further filtered to include only autosomal regions covered in at least 40 individuals, and removing sites with a minor allele frequency below 5% estimated from the GLs, resulting in the analyses being based on 603,519 variant sites. The highest likelihood solutions can be seen in Fig. 2c.
Principal Component Analysis
Assignment of individuals to ecotype, and structuring among ecotypes, was further investigated using Principal Component Analysis (PCA), implemented in the ngsTools suite54 taking genotype uncertainty into account55. Briefly, the covariance matrix between individuals, computed as proposed in ref. 56, is approximated by weighting each genotype for its posterior probability, the latter computed using ANGSD as described in ref. 18. The eigenvectors from the covariance matrix were generated with the R function ‘eigen’, and significance was determined with a Tracy–Widom test to evaluate the statistical significance of each principal component identified by the PCA. The PCA was plotted using an in-house R script (available at https://github.com/mfumagalli/ ngsPopGen/ tree/master/scripts).
Tests for ancestral admixture
The genetic relationships among ecotypes were further reconstructed in the form of a maximum likelihood graph representing the degree of genetic drift generated using TreeMix24. TreeMix estimates a bifurcating maximum likelihood tree using population allele frequency data and estimates genetic drift among populations using a Gaussian approximation. The branches of this tree represent the relationship between populations based on the majority of alleles. Migration edges are then fitted between populations that are a poor fit to the tree model, and in which the exchange of alleles is inferred. The addition of migration edges between branches is undertaken in stepwise iterations to maximize the likelihood, until no further increase in statistical significance is achieved. The directionality of gene flow along migration edges is inferred from asymmetries in a covariance matrix of allele frequencies relative to an ancestral population as implied from the maximum likelihood tree. We ran TreeMix fixing the transient ecotype as the root, with blocks of 1,000 SNPs (corresponding to approximately several hundred kilobases) to account for linkage disequilibrium among sites. The graphs are presented in Fig. 2a and Supplementary Fig. 5. This method does require genotype calls as an input, and is therefore susceptible to errors associated with genotype calling from low-coverage sequencing data. However, since TreeMix works on population allele frequencies, not genotypes, it was possible to determine the frequencies of the most common alleles with high confidence. The topology is comparable to output from other approaches applied here that do account for genotype uncertainty, providing confidence in the result.
Estimating population genomic metrics
Measures of genetic differentiation, divergence and diversity were estimated using methods specifically designed for low-coverage sequencing data. Using a Maximum-Likelihood-based approach previously proposed18 and using the bottlenose dolphin genome to determine the ancestral state of each site, the unfolded SFS was estimated jointly for all individuals within a population for sites sequenced in five or more individuals in each population. Using this ML estimate of the SFS as a prior in an Empirical Bayes approach, the posterior probability of all possible allele frequencies at each site was computed18. For these quantities, the expectations of the number of variable sites and fixed differences between lineages were estimated as the sum across sites of the probability of each site to be variable as previously proposed57. Finally, the posterior expectation of the sample allele frequencies was calculated as the basis for further analysis of genetic variation within and between lineages.
FST was estimated with a method-of-moments estimator58 based on both the maximum likelihood estimate of the 2D-SFS55 and the sample allele frequency posterior probabilities of the 2D-SFS55. The two estimates were highly correlated (Pearson’s correlation coefficient: r2>0.96) for all pairwise comparisons. However, from inspection of the data, the FST estimates generated from sample allele frequency posterior probabilities provided a more accurate estimation of fixed differences between populations. The likelihood-based method tends to flatten the FST peaks compared with the posterior probability method. This can result in masking of FST peaks with increasing genome-wide FST when using the likelihood-based method. Therefore, the posterior likelihood estimates are presented in the Manhattan plots of 50-kb sliding windows (Fig. 4a), with further filtering to only include windows for which >10 kb was covered by at least five individuals per population. We estimated the probability of a site being variable (Pvar). We tried different Pvar cutoffs and counted the number of variable sites at each Pvar. We then decided on Pvar=1, as the number of variable sites matched our expectations from estimates of diversity (π) from the two high-coverage genomes. We further checked by comparing FST estimates from a recently published whole-genome data set of carrion crows (Corvis corone) and hooded crows (C. cornix)59 downsampled to low coverage. By only considering sites with a Pvar of 1, we obtained FST estimates comparable to the values obtained from the 7–28 × sequences using GATK. Population-specific allele frequencies were estimated with the ancestral state fixed and the derived allele weighted by the probabilities of the three possible states. The allele frequencies were estimated based on GLs. To assess the robustness of our per-site FST value estimates, we evaluated how well the values correlated with those estimated from RAD-seq data generated in a previous study19. We accessed the SNP data in a VCF file format generated by the RAD-seq study from Dryad (doi:10.5061/dryad.qk22t) and used VCFtools to calculate per-site FST, using sites called at >20 × coverage, between 43 individuals of the resident ecotype and 37 individuals of the transient ecotype (Supplementary Fig. 2a). We then performed 1,000 replicates, randomly subsampling with replacement of 10 individuals from each ecotype. The correlation between FST estimates from these random subsamples ranged from 0.6861 to 0.9372. We found the correlations between estimates of FST from the RAD-seq data and those from our whole genome sequencing (WGS) data ranged from 0.5475 to 0.7140 (Supplementary Fig. 2b). The significant correlation in estimates of FST between two different methods using different individuals suggests that these estimates are reliable.
The average number of nucleotide substitutions Dxy was then calculated as the mean of p1*q2+p2*q1, where p1 is the allele frequency of allele 1 in population 1 and p2 in population 2, q1 is the allele frequency of allele 2 in population 1 and q2 in population 2. Sites that were not variable in any of the populations were assigned allele frequencies of zero.
Applying the probabilistic model implemented in ANGSD, we estimated the unfolded SFS in steps of 50-, 100- and 200-kb windows using default parameters and GLs based on the SAMtools error model. From the SFS we derived nucleotide diversity π (Supplementary Table 9). Estimates of nucleotide diversity can be influenced by differences in sequencing coverage and sequencing error. However, it has been shown that, using an empirical Bayes approach, implemented in ANGSD, the uncertainties in low-depth data can be overcome to obtain estimates that are similar to those obtained from data sets in which the genotypes are known with certainty60. Multiple checks were performed to ensure that estimates of π were not an artefact of the data-filtering. Comparable estimates of π were obtained using the method implemented in ANGSD for a single 20 × coverage ‘resident’ genome (π=0.0009) when it was randomly downsampled to 2 × coverage (π=0.0008). Nucleotide diversity was estimated for sites covered by at least five individuals in each population in windows of size 50, 100 and 200 kb (Supplementary Table 9).
PSMC analysis21 was performed on the high-coverage diploid autosomal genome sequences of two individuals to investigate changes in effective population (Ne) size. The PSMC model estimates the TMRCA of segmental blocks of the genome and uses information from the rates of the coalescent events to infer Ne at a given time, thereby providing a direct estimate of the past demographic changes of a population. The method has been validated by its successful reconstructions of demographic histories using simulated data and genome sequences from modern human populations21.
A more detailed account of PSMC analyses, including our reanalyses and reinterpretation of a previously published analysis of low-medium coverage genomes17, is presented in the Supplementary Figures and Notes. A PSMC plot of the Atlantic genome subsampled to 20 × and a North Pacific resident genome (20 × ) was estimated, scaled to the autosomal mutation rate of 2.34 × 10−8 substitutions per nucleotide per generation20 and is presented in Fig. 3a. A PSMC plot of the autosomal regions of the North Atlantic female killer whale at a coverage of 50 × was scaled to the autosomal mutation rate (μA) of 2.34 × 10−8 substitutions per nucleotide per generation20, as used in our estimation of TMRCA (see above) and was compared with a plot of the diploid X-chromosome, scaled to real time as per ref. 21 in which the neutral mutation rate of the X-chromosome was derived as μX=μA[2(2+α)]/[3(1+α)], assuming a ratio of male-to-female mutation rate of α=2 (ref. 61). This gave us an estimated μX=2.08 × 10−8 substitutions per nucleotide per generation. The plot is presented in Fig. 3b. We also re-estimated the PSMC plot for the X-chromosome using different mutation rates to investigate which rate would produce PSMC plots with inferred concurrent declines in Ne in autosomes and X-chromosome. We found that μX=1.00 × 10−8 substitutions per nucleotide per generation would be needed to synchronize the inferred demographic changes in these two markers (Supplementary Fig. 17). This would require the male-to-female mutation rate (α) to be orders of magnitude higher, making it seemingly biologically unrealistic.
To reconstruct ancestral demography in the ecotypes for which we did not have a high-coverage genome, we applied a method that uses the SFS from population genomic data to infer ancestral population size changes33. The Stairway Plot method first uses SFS from population genomic sequence data to estimate a series of population mutation rates (that is, θ=4Neμ, where μ is the mutation rate per generation and Ne is the effective population size), assuming a flexible multi-epoch demographic model. Changes in effective population size through time are then estimated based on the estimations of θ. As input data, we transformed the probability estimates of our site frequency spectra into SNP counts. We first ran the method on the resident ecotype to compare with the demographic reconstruction suggested by the PSMC analysis on the high-coverage North Pacific individual (see above). Population structure is a notable confounding factor for inferring demographic history30,62. Consistent with this, we estimate broad confidence intervals of our estimates of Ne subsequent to the estimated bottlenecks in each ecotype, during a period overlapping with previous estimates of within-ecotype lineage splits22. As we have sampled individuals from multiple subpopulations of the resident and transient (and possibly the Antarctic types, although less in known about population structuring in Antarctic waters), we potentially skew the SFS towards low-frequency polymorphism, thereby mimicking the pattern generated by population expansion63. We therefore opted not to include SNP counts for singletons and doubletons, which are expected to have arisen recently within each ecotype and had no time to be shared throughout the population, as these may be biased by our sampling protocol and low-coverage sequencing data, and as our interest was in demographic change during population splits, which were estimated to have occurred over 10,000 years ago. Our focus is on the timing and extent of the bottlenecks within each ecotype, which all individuals within an ecotype coalesce back to and therefore pre-date within-ecotype population splits and substructuring and the emergence of derived singleton and doubletons.
The inference of demographic history from population genomic data by the Stairway Plot method provided a somewhat comparable result to reconstruction from a single 20 × genome by PSMC with respect to the time of onset and magnitude of a demographic decline inferred by both methods. The Stairway Plot method inferred a sudden drop in Ne, whereas a more gradual decline was inferred by PSMC, consistent with simulations showing that PSMC can infer abrupt changes in Ne as gradual changes21. The sudden change in estimated effective population size in the stairway plot is because of the method being based on a multi-epoch method, in which epochs coincide with coalescent events33. Therefore, the plot is not continuous, but rather it depicts discrete blocks of time (epochs). The number of epochs is determined by the number of individuals within each sample and the number of SNP bins used, that is, the number of possible coalescent events. The Stairway Plot method inferred a subsequent and rapid expansion, whereas PSMC did not infer an expansion within our cutoff point of 10,000 years ago, but did infer a more gradual expansion during the Holocene. It should be kept in mind that the PSMC plot is based on data from a single individual and so will track declines in Ne because of further founder effects as the resident ecotype continues to split into multiple discrete populations. In contrast, the Stairway Plot is based on population genomic data and will track the change in Ne across all the sampled resident populations after they have split.
The method was then applied to the site frequency spectra of the transient ecotype and type C. The results are shown in Fig. 3c, and in each case the Stairway Plot infers a sudden and dramatic demographic decline, consistent with previously inferred population split times followed by a demographic expansion. The timing of the decline of the Antarctic types overlapped because of the recent shared ancestry, and therefore only the plot for type C is shown in Fig. 3c for clarity.
Inferring putatively functional allele shifts because of selection
Shifts in allele frequencies can occur because of selection, but differences in allele frequencies can also accumulate between populations because of drift34,35. To infer shifts in allele frequencies potentially due to selection, we considered the 1% of SNPs with the highest FST values from each pairwise comparison between killer whale ecotypes. However, as FST is dependent on the underlying diversity of the locus, even extreme outlier loci can be due to genetic drift alone. We, therefore, additionally looked for over-representation of the top 1% outliers in different categories, for example, exons, 25-kb flanking regions (potential regulatory regions), introns and intergenic regions, at different FST bins using a χ2-test. Residuals are expected to be normally distributed and indicate statistical significance of over- and under-representation of specific categories. The significance threshold was subjected to Bonferroni correction. The top 100 outliers in exons from each pairwise comparison were then used for GO over-representation analysis64 to identify enrichment due to diet (in mammal- versus fish-eating ecotypes) and climate (in Antarctic versus Pacific).
To more robustly infer whether genetic changes in exons were associated with ecotype divergence due to selection, we applied the PBS40. The PBS has strong power to detect (even incomplete) selective sweeps over short divergence times5,40,65, making it relevant for the scenario we are investigating in this study. We therefore estimated the PBS for 50-kb sliding windows shifting in 10-kb increments (approximating to a window size of 10 SNPs) to detect regions of high differentiation potentially due to selective sweeps. We followed the approach of Yi et al.40, and used the classical transformation by Cavalli-Sforza66, T=−log(1−FST) to obtain estimates of the population divergence time T in units scaled by the population size. For each 50-kb window, we calculated this value between each pair of ecotypes. The length of the branch leading to the one ecotype since the divergence from a recent ancestor (for example, in the equation given the length of the branch to type B1 since diverging from types B2 and C) is then obtained as
These window-based PBS values represent the amount of allele frequency change at a given 50-kb genomic region in the history of this ecotype (since its divergence from the other two populations)40. To further narrow down the target of selection, the PBS was estimated for exons as per ref. 5 and compared with genome-wide values to identify whether they were in the top 99.9 percentile (for the branches to the resident, transient and ancestral Antarctic ecotypes) and the top 99.99 percentile for the shorter branches leading to each Antarctic ecotype from their most recent common shared ancestor. We further filtered outliers to just include exons that encoded non-synonymous amino-acid substitutions and searched the database GeneCards67 for functions of the encoded proteins to identify potential targets of natural selection due to ecological differences. Supplementary Table 10 contains a list of outlier loci and the corresponding PBS value for each branch. Supplementary Data 2 contain details of fixed non-synonymous changes in the exons of protein-coding genes, including the per-individual and per-population read counts for each allele.
Accession codes: All sequencing data are available in European Nucleotide Archive (ENA) under accession numbers ERS554424– ERS554471 (see Supplementary Table 1 for a full list of accession numbers and associated sample ID).
How to cite this article: Foote, A. D. et al. Genome-culture coevolution promotes rapid divergence of killer whale ecotypes. Nat. Commun. 7:11693 doi: 10.1038/ncomms11693 (2016).
The project was funded by a Marie Curie IEF ‘KWAF10’ and Lawksi Foundation grants to A.D.F.; European Research Council grant ERCStG-336536 to J.B.W.W.; a Danish National Research Foundation grant DNRF94 to M.T.P.G.; by Swiss SNSF Grant 31003A-143393 to L.E.; and by a short visit grant to A.D.F. from the ESF Research Networking Programme ConGenOmics. We acknowledge The Danish National High-Throughput DNA Sequencing Centre for sequencing the samples and, particularly, Andaine Seguin-Orlando, Lillian Petersen, Cecilie Demring Mortensen, Kim Magnussen and Ian Lissimore for technical support. Large-scale computational effort was made possible by the UPPMAX next-generation sequencing cluster and storage facility (UPPNEX), funded by the Knut and Alice Wallenberg Foundation and the Swedish National Infrastructure for Computing. Heng Li, Richard Durbin, Line Skotte, Anders Albrechtsen, Stephan Schiffels, Lounès Chikhi and Rasmus Nielsen provided useful feedback and discussions on the methods. We thank Willy Rodríguez for providing scripts for the ms simulations. Mike Ford and Kim Parsons provided useful feedback on an earlier draft of this manuscript. We thank Paul Fiedler for providing the map used in Fig. 1 and Uko Gorter for providing the illustrations of killer whale ecotypes, and Robert Pitman’s support and input throughout this project. We are grateful to M. Ford, K. Parsons, A. Burdin, M. Dahlheim, R. Pitman, the SWFSC Marine Mammal and Sea Turtle Research Collection, and to the International Whaling Commission for providing samples. We dedicate this work to the memory of Eva Saulitis, whose rich legacy includes an invaluable contribution to the foundations of our understanding of killer whale biology, without which this study would not have been possible.
D-statistic (ABBA-BABA test) values and standard error estimates for the inference of ancient admixture.
Details of fixed non-synonymous changes in the exons of protein coding genes, including the per-individual and per-population read count for each allele.
Pairwise genetic distance estimates and 100 bootstrap replicates.
Pairwise relatedness estimates
This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/