Complex patterns of sex-biased demography in canines

Abstract

The demographic history of dogs is complex, involving multiple bottlenecks, admixture events and artificial selection. However, existing genetic studies have not explored variance in the number of reproducing males and females, and whether it has changed across evolutionary time. While male-biased mating practices, such as male-biased migration and multiple paternity, have been observed in wolves, recent breeding practices could have led to female-biased mating patterns in breed dogs. For example, breed dogs are thought to have experienced a popular sire effect, where a small number of males father many offspring with a large number of females. Here we use genetic variation data to test how widespread sex-biased mating practices in canines are during different evolutionary time points. Using whole-genome sequence data from 33 dogs and wolves, we show that patterns of diversity on the X chromosome and autosomes are consistent with a higher number of reproducing males than females over ancient evolutionary history in both dogs and wolves, suggesting that mating practices did not change during early dog domestication. By contrast, since breed formation, we found evidence for a larger number of reproducing females than males in breed dogs, consistent with the popular sire effect. Our results confirm that canine demography has been complex, with opposing sex-biased processes occurring throughout their history. The signatures observed in genetic data are consistent with documented sex-biased mating practices in both the wild and domesticated populations, suggesting that these mating practices are pervasive.

1. Introduction

Dogs are the first animals known to be domesticated and have lived alongside humans and shared our environment ever since [1]. There is tremendous interest in understanding their genetics and evolutionary history [2–5]. Many studies have shown that dogs have a complex evolutionary history; they experienced a population size reduction (i.e. bottleneck) associated with domestication and additional breed-specific bottlenecks associated with breed formation during the Victorian era [6]. In addition to bottleneck events, dogs experienced admixture with wolves during the domestication process [7]. Lastly, dogs have experienced tremendous artificial selection for body size, coat colour and other phenotypes [3,8–10]. There is disagreement about the process of domestication, including when, where and how many times dogs were domesticated [2,11–15]. However, despite the extensive work on understanding dog demographic history, existing studies have not explored the sex-specific population history of males and females across dog domestication. Departures from an equal number of reproducing males and females are called sex-biased demographic processes, and leave signatures in the genome (reviewed in Wilson Sayres [16]). Previous ecological and field studies suggested that mating practices have been sex biased in canines. In the wild populations, Yellowstone male wolves sometimes migrate to an existing wolf pack to mate with the alpha female when the alpha male dies [17]. The migration into an existing wolf pack is therefore male biased. An additional source of male-biased migration may come from male wolves called ‘Casanova wolves’. These wolves leave their natal packs and visit a nearby wolf pack around mating season to mate with subordinate females [18]. Lastly, there is evidence of multiple paternity in Ethiopian wolves and foxes [19,20]. In domesticated populations, it is thought that more females contributed to breed formation than males, indicating female-biased processes [21]. In addition, recent reproductive practices, such as the popular sire effect, which involves a small number of males reproducing with a large number of females can lead to female-biased demography [22]. Despite these observations of mating practices suggesting the numbers of reproducing males and females has been unequal during canid evolution, it is unclear how pervasive these processes are, and which have had the dominant effect on shaping patterns of genetic diversity.

To test how widespread sex-biased demography has been throughout canid evolution, we inferred the effective population size on the X chromosome and the autosomes. The ratio of the effective population size on the X chromosome to that of the autosomes was termed Q in Emery et al., and we will use this notation throughout [23]. In male-heterogametic sex-determining systems (XX/XY) with equal numbers of reproducing males and females, there are three copies of the X chromosome for every four copies of each autosome. Therefore, in a constant size population without any natural selection or demographic bias (sex-shared or sex biased), Q is expected to be 0.75 (reviewed in Webster & Wilson Sayres [24]). Specifically, Q = N_X/N_A ≅ 0.75. Deviations from this expected ratio could be indicative of sex-biased processes. Q less than 0.75 is consistent with fewer copies of the X chromosome than expected, suggesting a larger number of reproducing males than reproducing females, indicative of male-biased processes. Q greater than 0.75 is consistent with more copies of the X chromosome than expected, suggesting a larger number of reproducing females than reproducing males, indicative of female-biased processes. If Q is inferred using the ratio of genetic diversity on the X chromosome to the autosomes (Q_π = π_X/π_A), deviations from 0.75 could also be owing to demographic events, rather than sex-biased processes. Pool & Nielsen [25] showed that Q_π less than 0.75 could occur when there is an equal number of reproducing males and females if there has been a population bottleneck because the X chromosome loses more of its genetic diversity than the autosomes.

Studies comparing measures of genetic diversity between the X chromosome and autosomes have resulted in many insights into the evolutionary history of humans. Hammer et al. [26] computed Q by fitting a model of demographic history to the number of single-nucleotide polymorphisms on the X chromosome and autosomes. They found that Q is greater than 0.75 in all human populations examined, suggesting female-biased processes that have led to more reproducing females than males during human evolutionary history [26]. Later, Keinan et al. [27] computed Q by calculating the ratio in fixation index, F_ST, between the X chromosome and the autosomes: $Q_{\mathrm{FST}}=\ln (1-2F_{\mathrm{ST}}^A)/\ln (1-2F_{\mathrm{ST}}^X)$. They found that Q_FST is less than 0.75 only when comparing a non-African population to an African population [27]. They interpreted this result to suggest that there was a male-biased migration out of Africa, where there were more reproducing males than females. Even though these two studies came to different conclusions regarding the sex ratio in human history, a later study reconciled these seemingly disparate findings by demonstrating that Q can detect bias in sex ratios at different timescales, depending on whether it is calculated from genetic diversity (Q_π) or the fixation index (Q_FST) [23]. Specifically, Q_π can detect sex bias in ancient timescales (i.e. more in the past, before the populations split from each other), whereas Q_FST detects sex-biased demography on recent timescales, after the populations split from each other [23]. The reason for this is that Q_π uses the number of average pairwise differences between sequences, and as such, it is heavily influenced by common variants that have occurred during the entire history of the population. Q_FST on the other hand, measures the amount of genetic drift occurring on the X chromosome as compared to the autosomes only after the two populations split from each other. Emery et al. [23] reconciled results from Hammer et al. [26] and Keinan et al. [27] by showing that evolutionary processes within human history could be consistent with an earlier female bias followed by a male bias during the migration of some humans out of Africa [23]. Additionally, direct comparisons of the two studies were complicated by linked selection on the X chromosome [28,29]. Beyond humans, comparing the genetic diversity between the X chromosome and autosomes has also been used to study sex-biased processes in other species [16]. For example, Chen et al. [30] examined X/A diversity ratio in sheep and found evidence for significantly lower genetic diversity on the X chromosome as compared to autosomes, indicative of sex-biased processes [30].

Given how examining patterns of genetic diversity on the X chromosome and the autosomes has facilitated our understanding of sex-biased demography in other species, and what has been observed regarding sex-biased mating practices in canines, we set out to test how widespread these mating practices are throughout different time points during canine evolutionary history. We characterized patterns of genetic variation across whole-genome sequences of 21 dogs and 12 wolves. We find that Q on ancient timescales is less than 0.75 for both taxa, suggesting a greater number of reproducing males than females. In addition, using the estimator of the effective sex ratio based on the fixation index, we showed that while the demographic history in wolves has remained male biased in recent history, the demographic history in dogs has changed from male biased in the ancient timescale to female biased in recent times. Our work adds to the current understanding of canine demographic history and suggests the need to incorporate sex-biased demography in future studies.

2. Results

(a). Description of the data

We analysed 33 female canid whole genomes that include four German shepherds, five Tibetan mastiffs, 12 dog individuals from a variety of breeds, six Arctic wolves and six grey wolves (see the electronic supplementary material, table S1). Fastq files for German shepherd and Tibetan mastiff data were downloaded from NCBI SRA [31]. We combined 12 high coverage (greater than 15X) whole-genome sequences of female dogs from multiple breeds [32] because we were interested in how results differ between using a group of one breed versus using a group consisting of multiple breeds. We named this pooled group the ‘pooled breed dogs’. The Arctic wolves were located in northern Canada (north of the Arctic circle) [33]. We also used high coverage (greater than 15X) whole-genome sequences of female grey wolves [32]. Because these grey wolves originated from Europe, Asia and Yellowstone, we named this population the ‘pooled grey wolves’. Details about coverage and accession numbers for the individuals in this study are summarized in the electronic supplementary material, table S1. We focused on these particular populations because of the availability of at least four female individuals per sample with high coverage (greater than 15X) genome sequence data. Because males have one X chromosome while females have two X chromosomes, we only analysed female samples in this study to avoid differences in genotype calling on the X chromosome between males and females.

To understand population structure in our data, we performed a principal component (PC) analysis on the autosomes and the X chromosome separately using SNPRelate [34]. On both the autosomes and the X chromosome, we observed that PC1 separates the dogs from the wolves while PC2 separates out the Arctic wolves from the grey wolves (see the electronic supplementary material, figure S1A). When we subset the data to include only the dog populations, among dogs we observed that PC1 separates the three different dog groups from each other (see the electronic supplementary material, figure S1B). As seen in other studies [35], the Basenji does not cluster with other breed dogs, reflecting its ancient origins (see the electronic supplementary material, figure S1B).

(b). Estimating the effective sex ratio based on genetic diversity

Previous work has shown that dogs experience male mutation bias, where the mutation rate is higher in males compared to females owing to more germline cell divisions in males at reproduction [36–38]. Using dog–cat divergence, we observed that the level of male mutation bias is around 2, which is consistent with previous reports [37,38] (see the electronic supplementary material, table S2). Therefore, we controlled for male mutation bias in all estimates of genetic variation by normalizing autosomal and X chromosome diversity by dog–cat divergence in the corresponding regions. To control for differences in number of sites where mutations would be directly affected by natural selection on the X chromosome and the autosomes, we used regions of the genome in which mutations would be putatively neutral by removing genic and conserved sites.

To understand whether any evolutionary process has been sex biased over ancient timescales, we first calculated genetic diversity, π, on the autosomes and on the X chromosomes and normalized genetic diversity by dog–cat divergence to correct for mutation rate variation (see the electronic supplementary material, tables S3 and S4). We computed Q_π by taking the ratio of π on the X chromosome and autosomes. We found that in both dog and wolf populations, Q_π is significantly less than 0.75 (figure 1, no centimorgan (cM) cutoff).

Figure 1.

X-linked and autosomal genetic diversity across canids. Genetic diversity measured as the average pairwise differences between sequences (π) corrected for mutation rate variation using divergence on the X chromosome and autosomes. Q denotes the ratio of π on the X chromosome to that of the autosomes. The horizontal red line denotes the null expectation of 0.75. Bins along the x-axis denote different filtering based on genetic distances from genes. Error bars denote 95% confidence intervals obtained through bootstrapping. cM, centimorgan. (Online version in colour.)

Q_π of less than 0.75 could occur owing to the effect of natural selection on linked neutral sites. Specifically, natural selection could have reduced diversity in linked neutral regions on the X chromosome more than on the autosomes, as seen in humans [27–29]. Further, it is possible that there is more constraint on non-coding regions near genes on the X chromosome than on the autosomes [39]. To measure how neutral diversity is affected by linked selection, we compared diversity on the X chromosome and autosomes in regions near genes versus putatively unconstrained regions 0.4 cM away from the nearest gene. Diversity increased more with increasing distance from genes on the X chromosome than on the autosomes, consistent with natural selection reducing diversity more on the X chromosome than on the autosomes near genes (see the electronic supplementary material, table S5).

To test whether stronger linked selection acting on the X chromosome relative to the autosomes could cause Q_π to be less than 0.75, we expanded our filtering criteria to remove sites that are near genes, defined by genetic distance. Because we did not know a priori what the minimum genetic distance would be required to obtain sites that are not affected by selection, we included several thresholds. We removed sites whose genetic distance to the nearest genes is less than 0.2, 0.4, 0.6, 0.8 and 1 cM. We observed that even after removing sites whose genetic distance to the nearest genes are less than 1 cM, Q_π is still less than the expected 0.75 in both dog and wolf populations, except for the German shepherd (figure 1). In the German shepherd, when using the thresholds of 0.8 and 1 cM, Q_π approaches 0.75. However, because there are significantly fewer sites and variants left after removing sites whose genetic distance to the nearest genes is less than 0.8 or 1 cM, we could not exclude the possibility that we are underpowered to detect any signal in the data (see the electronic supplementary material, table S6). Nonetheless, these results suggest that while linked selection may partially account for Q_π of less than 0.75, especially in the German shepherd, linked selection by itself cannot explain why Q_π is less than 0.75 across all dog and wolf populations. The fact that Q_π is less than 0.75 in both dog and wolf populations suggests there has been male-biased sex ratios in both dogs and wolves over ancient (i.e. prior to the time they split from each other) evolutionary timescales. However, in principle, Q_π of less than 0.75 could be explained by population bottlenecks even with an equal number of breeding males and females [25].

(c). Inference of sex-biased demographic processes under population genetic models

To test whether population bottlenecks can explain the reduction in diversity on the X chromosome, we fitted a demographic model that includes a bottleneck using the autosomal site frequency spectrum (SFS; see the electronic supplementary material, figure S2) and asked whether the best-fitting demographic model on the autosomes could also account for the level of diversity on the X chromosome when Q equals 0.75. If a demographic model including a bottleneck by itself can generate a Q_π of less than 0.75, we would expect that scaling the population size of the X chromosome to be three-quarters that of the autosomes should result in a Q_π comparable to the empirical data. Additionally, we then employed a composite likelihood framework to directly infer the N_X/N_A ratio from the SFS while accounting for the complex non-equilibrium demography.

First, we fitted a demographic model that includes a bottleneck using the SFS on the autosomes using fastsimcoal2 [40] for each population considering regions of greater than 0.4, 0.6, 0.8 and 1 cM from genes. We reasoned that we would not be able to exclude the role of selection when not removing sites near genes or using too small a threshold (i.e. 0.2 cM). We also corrected for male mutation bias using mutation rates that we inferred from dog–cat divergence in the same windows (see the electronic supplementary material, table S2). The inferred demographic parameters that resulted in the best likelihood of the data are presented in the electronic supplementary material, table S7. To test whether the inferred demographic parameters can recapitulate the autosomal data, we used fastsimcoal2 to generate the expected SFSs. In all populations except the German shepherds, across all thresholds examined, we observed that the SFSs generated using the inferred demographic parameters visually match with the empirical autosomal SFSs (see the electronic supplementary material, figure S3). The differences in log-likelihood between the simulated SFSs and the empirical SFSs are also small (see the electronic supplementary material, table S8), confirming our visual inspection of the fit of the demographic models. In addition, autosomal genetic diversity (π) computed from the demographic model is comparable to the empirical estimates of π (see the electronic supplementary material, figure S4). Thus, these lines of evidence demonstrate that the inferred demographic parameters can recapitulate the empirical data on the autosomes, except for the more stringent filtering on the German shepherd (see Discussion).

We then checked whether the inferred demographic parameters fit the SFS for the X chromosome. To account for the differences in population size between the X chromosome and the autosomes, we adjusted the population size on the X chromosome by a constant value which we called C, where N_X = CN_A. If a bottleneck by itself without any sex-biased demography can generate a Q_π of less than 0.75, we expected that using a C value of 0.75 would recapitulate the empirical data. If a bottleneck model by itself is not sufficient to generate a Q_π of less than 0.75, and sex-biased processes need to be invoked, we expected that rescaling the population size on the X chromosome to be three-quarters of the population size on the autosomes would not fit well. Rather, a different value of C would yield a better fit.

To assess whether a null C value of 0.75 or a different C value yielded a better fit to the empirical SFSs on the X chromosome, we searched over a grid of C values. We found the maximum-likelihood value of C for each population and filtering threshold. To do this, for each C on a grid of C values, we first calculated the population size on the X chromosome, which is N_X = CN_A. We then used fastsimcoal2 to simulate an SFS and assess the fit by comparing the Poisson log-likelihood to the SFS on the X chromosome. For each population and for each threshold, we found a set of C values that maximizes the likelihood of the data (table 1; see the electronic supplementary material, figure S5 and table S9).

Table 1.

Likelihood ratio tests comparing models of sex-biased demography in canid populations. (Likelihood ratio tests of the amount of sex-biased demography are shown when removing any sites whose genetic distance to the nearest genes is less than 0.4 cM.)

population	C	log-likelihood	likelihood ratio test statistic	p-value
German shepherds	null (C = 0.75)	7460.957	34.779	3.69 × 10−9
	best (C = 0.68)	7478.347
Tibetan mastiffs	null (C = 0.75)	16050.84	208.972	2.30 × 10−47
	best (C = 0.61)	16155.32
pooled breed dogs	null (C = 0.75)	16875.26	249.627	3.13 × 10−56
	best (C = 0.61)	17000.07
Arctic wolves	null (C = 0.75)	23035.46	133.341	7.62 × 10−31
	best (C = 0.65)	23102.13
pooled grey wolves	null (C = 0.75)	30767.69	188.719	6.05 × 10−43
	best (C = 0.64)	30862.05

With the exception of the German shepherd at the most stringent filtering thresholds (greater than 0.8 cM and greater than 1 cM), we inferred that C is less than 0.75 for all populations and filtering thresholds. When using a filtering threshold of 0.4 cM from genes, we found that C ranges from 0.61 to 0.68. The full model, where we inferred C for each comparison, fits the observed X chromosome SFS significantly better than a model where C is constrained to be 0.75 (likelihood ratio test statistic greater than 30, 1 d.f., p-value < 10⁻⁸; table 1). Further, the null C value of 0.75 does not visually fit the SFSs on the X chromosome (see the electronic supplementary material, figure S6, blue bars), suggesting that we can reject an equal number of reproducing males and females, even in the presence of a bottleneck. Third, we observed that diversity on the X chromosome from simulating with a null C value of 0.75 overestimated the empirical X chromosome diversity (see the electronic supplementary material, figure S7, blue bars). These results suggest that a model including both a bottleneck and a male-bias sex ratio can generate Q_π of less than 0.75 and recapitulate the observed SFSs and genetic diversity. Only in the German shepherd population when using the most stringent threshold (greater than 0.8 cM and greater than 1 cM), can a demographic history including a bottleneck by itself match the empirical patterns.

(d). Female-biased sex ratio within dogs in recent history

Since estimates of sex ratios from levels of genetic diversity are sensitive to ancient sex-biased processes (prior to or immediately after the split between two species), we wanted to determine whether the pattern of male-biased contributions remained constant throughout the evolutionary history of canines [23]. To study sex-biased demography on recent timescales, we computed Q_FST for each pair of populations. In the dog to dog comparison, we computed Q_FST between German shepherds and Tibetan mastiffs, between German shepherds and pooled breed dogs, and between Tibetan mastiffs and pooled breed dogs (see the electronic supplementary material, table S10). We observed that Q_FST is greater than 0.75 for all three pairs and across all thresholds, suggesting a female-biased sex ratio within the dog populations in recent history (figure 2 and electronic supplementary material, figure S8 and table S11). This is consistent with fewer reproducing males than females in the population since the formation of different dog breeds. In the wolf to wolf comparison, we computed Q_FST between Arctic wolves and pooled grey wolves. In contrast to the breed dogs, we found that Q_FST is less than 0.75 when using the thresholds of greater than 0.4 cM and greater than 0.6 cM, suggesting that a male-biased sex ratio has been maintained within the wolf populations in recent history (figure 2 and the electronic supplementary material, figure S8). However, we noted that when using a more stringent threshold (greater than 0.8 cM or greater than 1 cM), Q_FST within wolves approaches 0.75 or greater than 0.75 (see the electronic supplementary material, figure S8). We could not exclude the possibility that we are unable to detect a true signal in the data due to significantly fewer sites and variants left after the more stringent filtering (see the electronic supplementary material, table S6). Overall, these results indicate that while the process within wolves has probably maintained a male bias from ancient to recent history, the process within dogs has changed to a female bias, potentially because of breeding practices that have led to female-biased processes such as the popular sire effect.

Figure 2.

Sex-biased demography on recent timescales. Estimates of the sex ratio for a pair of populations computed using F_ST using a threshold of greater than 0.6 cM to remove linked neutral sites are shown. The horizontal red line denotes the null expectation of 0.75. Error bars denote 95% confidence intervals obtained through bootstrapping. GS (German shepherds), TM (Tibetan mastiffs), BD (pooled breed dogs), AW (Arctic wolves) and GW (pooled grey wolf). (Online version in colour.)

(e). Approximate Bayesian computation framework to infer sex-biased demography

To better quantify how sex-biased demographic processes have changed at different times, we employed an approximate Bayesian computation (ABC framework) as implemented in ABCtoolbox [41]. For this analysis, we focused on regions more than 0.6 cM from genes. We inferred demographic parameters using the summary statistics π and F_ST on the autosomes in a two-population split demographic model (figure 3a; see the electronic supplementary material, table S12). For dogs, we analysed the German shepherds and the Tibetan mastiffs. For wolves, we compared the Arctic wolves and the pooled grey wolves. We first inferred the demographic parameters on the autosomes, specifically the current size of population 1 since the split (NPOP1) and of population 2 (NPOP2) (see parameter estimates in the electronic supplementary material, table S13). The inferred model fits the autosomal summary statistics well as shown by posterior predictive simulations (see the electronic supplementary material, figures S9 and S10, table S14). We then used the summary statistics from the X chromosome to infer the degree of sex-biased demography before the populations split from each other (C_ancient; figure 3b) and the degree of sex bias after the two populations have split from each other (C_recent; figure 3b). We performed three replicates of our inference procedure to ensure consistency of the results. Within dogs, across all three replicates, we found that the entire posterior distribution of C_ancient is less than 0.75. For example, for one of the replicates, median C_ancient = 0.380, 95% credible interval: 0.353–0.406 (see the electronic supplementary material, table 13 for median values for the other replicates). This result supports our empirical findings using Q_π as well as our inferences of C from the SFS that on ancient timescales (i.e. prior to the formation of different dog breeds) there was an excess of breeding males compared to females, decreasing the effective population size on the X chromosome. When analysing the two dog populations, across all three replicates, C_recent > 0.75. For example, for one replicate, the median of the posterior distribution of C_recent = 1.006 (95% credible interval: 0.859–1.152 (figure 3c; electronic supplementary material, table S13). Our finding of C_recent > 0.75 supports our Q_FST analysis indicating that in recent times, after the two dog breeds split from each other, there was a decrease in the number of breeding males compared to breeding females. On the other hand, within wolves, C_ancient < 0.75 (for one of the replicate, C_ancient = 0.536, 95% credible interval 0.506–0.599), and the credible interval of C_recent spanned 0.75 (figure 3d; electronic supplementary material, table S13). These findings are again concordant with our inferences using Q_π, Q_FST and the likelihood-based inferences of C using the SFS.

Figure 3.

Inferring sex-biased demography at different timescales under an approximate Bayesian framework. (a) A two-population split model was used to infer demographic parameters. We used the summary statistics on the autosomes to infer the current population sizes, NPOP1 and NPOP2. (b) Model used to infer the sex ratio at different timescales. We scaled the ancient population size by C_ancient. We scaled the current population sizes by C_recent. We inferred the posterior distributions of C_ancient and C_recent using the summary statistics on the X chromosome. (c) Prior and posterior distributions of C_ancient and C_recent for two dog populations, German shepherds and Tibetan mastiffs. (d) Prior and posterior distributions of C_ancient and C_recent for two wolf populations, Arctic wolves and grey wolves. Because the results for all three replicates are similar, we show the results for replicate 2. See the electronic supplementary material, table S13 for results for the other two replicates. (Online version in colour.)

3. Discussion

In this study, we used two different statistics and two different model-based inference methods to estimate the ratio of reproducing males to females in canines and found that the demographic history of dogs and wolves has been sex biased, but not always in the same direction. Estimating the sex ratio based on the levels of genetic diversity (Q_π) from the X chromosome and autosomes showed a male-biased sex ratio in both dogs and wolves on an ancient timescale (i.e. prior to divergence between two species), which cannot be explained by linked selection or a population size reduction on its own (figure 1, table 1 and electronic supplementary material, figure S5). Instead, in both dogs and wolves, there has been a larger number of reproducing males than females. Our inferences of ancient mating practices are consistent with mating practices observed in wolf packs today. For example, the alpha male and female are the dominant reproducers in wolf packs, but subdominant reproduction is common and may involve multiple fathers for a single litter [17]. Multiple paternity is a unique aspect of canid reproduction and may help drive a male bias in reproduction, as offspring of a single litter can only have a one mother, but may have multiple fathers and litter size may be as large as 16 individuals [42]. In addition, wolves migrating to existing wolf packs are predominantly male biased [17]. Further, ‘Casanova wolves’ who stay near a wolf pack during mating season to mate with the non-alpha females could also cause male-biased mating patterns [18]. However, since dog breed formation, both Q_FST and the ABC inference show that there is an excess of breeding females relative to males. These results are consistent with breeding practices in modern dogs. While multiple paternity and male-biased migration probably occurred in early dogs, but under more recent controlled breeding, valuable sires would be the only father of a litter. Hence the controlled nature of breeding in modern dog breeds, and the focus on a subset of ‘popular’ sires could drive the female bias in reproduction. The population sire effect also reduces the effective size of breeds and effects such as inbreeding further skew evolution in modern breeds.

Some limitations in this study provide avenues for future work. First, our study was limited by the availability of high coverage (greater than 15X coverage) whole-genome sequences of female individuals at the time of analysis. Future studies could use more female individuals and a variety of populations to understand whether there are differences in sex-biased processes between breeds. Second, it is possible that there have been shifting selective pressures in dogs compared to wolves. For example, it is possible that mutations which are more strongly deleterious in wolves might be rendered less deleterious in dogs, owing to decreasing selective pressures relating to domestication [3]. Alternatively, there may have been more selective sweeps in one species or the other that may impact the X differently from the autosomes. However, selection is unlikely to affect our conclusions as we focus our analyses on non-coding and non-genic regions greater than 0.6 cM distant from conserved or genic regions. Although linked selection may extend greater than this distance in dogs owing to strong artificial selection, we find the ancient male bias in both dogs and wolves. The fact that wolves have not experienced strong artificial selection suggests that this pattern is not driven by selection. Further, there likely has been admixture between dogs and wolves at various points throughout their history [2,43–45]. Our demographic inferences are based on models assuming no gene flow between these populations. However, our conclusions are likely to be robust to some level of admixture. A previous simulation study has suggested that recent admixture is unlikely to bias inferences of demography in the ancestral population when using the SFS [46]. While it is difficult to assess how admixture could affect inferences of sex-biased demography, the fact we find that $N_X/N_A<0.75$ during ancient history in both dogs and wolves suggests that admixture would not affect this pattern. Future studies could examine whether processes such as admixture with wolves or introgression has been sex biased.

Finally, in all populations our demographic inference only included one bottleneck per population. Though the models fit the SFS well (see the electronic supplementary material, figure S3), the single bottleneck model is probably averaging over bottlenecks associated with domestication and breed formation in dogs. Future work could extend our modelling framework by including more complex demographic scenarios such as multiple population bottlenecks and migration events to better capture the autosomal data, especially the German shepherds. In the German shepherds, using filtering thresholds of greater than 0.6, 0.8 and 1 cM, there is a deficit in the number of doubletons in the empirical data (see the electronic supplementary material, figure S3) compared to what is predicted by the best-fitting model, suggesting that the demographic model including just the bottleneck may not be sufficient. We acknowledge that the patterns in the German shepherds are different from the other populations examined in this study and is worth a more in-depth investigation to obtain the demographic parameters that can recapitulate the empirical SFSs. As such, future studies could improve on this study by including more German shepherd individuals and exploring other demographic scenarios that could better fit the SFS.

In conclusion, our results add to the growing literature on the complex demographic history of dogs [3,4]. In addition to multiple episodes of bottleneck and admixture events, we now present evidence for sex-biased demographic processes. Furthermore, we provide evidence that sex-biased processes within dogs have changed throughout evolution, switching from a male bias in ancient timescales to a female bias in recent timescales, reflecting how modern breeding practices influence the sex ratio.

4. Material and methods

(a). Whole-genome sequence processing

We used Genome Analysis Toolkit (GATK) version 3 for variant discovery and followed the documentation for best practices [47–49]. Details on our procedures and filtering can be found in the electronic supplementary material and scripts used for processing whole-genome sequencing data can be found at https://github.com/tnphung/NGS_pipeline.

(b). Population genetic analyses

We computed genetic diversity, π, defined as the average number of differences between pairs of sequences [50]:

$$\pi ={\displaystyle \frac{n}{n-1}\sum _i^{\mathrm{all}\hspace{0.17em}\mathrm{sites}}{p}_i(1-p_i),}$$

where p_i is the allele frequency and n is the number of alleles. To compute Q_FST, we followed the approach described previously [23,27] and computed Weir & Cockerham's F_ST for each pair of populations using the SNPRelate package implemented in R [34].

We inferred demographic parameters from the autosomal data (SFSs on the autosomes) using a maximum-likelihood framework as implemented in fastsimcoal2 version 2.6 [40]. We specified a bottleneck demographic model and inferred four parameters: N_ANC which is the population size in the ancestral population, N_BOT which is the population size during the bottleneck, N_CUR which is the population size in the current day, and T_BOT which is the duration between the end of the bottleneck and current day (see the electronic supplementary material, figure S2). To account for differences in population size between the X chromosome and autosomes, we scaled the population size on the X chromosome to that on the autosomes by a constant factor we called C, where N_X = CN_A. To find the maximum-likelihood estimate of C, we searched over a grid for values of C, including 0.75, to find a value that resulted in the highest likelihood. See the electronic supplementary material for more details on the inference procedures.

Data accessibility

All scripts can be found at https://github.com/tnphung/SexBiased. SRA numbers for fastq files for published genomes are listed in the electronic supplementary material, table S1. Newly sequenced Arctic wolf reads have been deposited into the SRA with accession numbers: SRS4647039, SRS4647041, SRS4647038, SRS4647040, SRS4647036, SRS4647037. Data (filtered VCF and BED files representing putatively neutral regions) available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.jd820r4 [51].

Authors' contributions

T.N.P., R.K.W., M.A.W. and K.E.L. designed research. T.N.P. analysed data. T.N.P., M.A.W. and K.E.L. wrote the manuscript with significant input from R.K.W. M.A.W. and K.E.L. jointly supervised this work.

Competing interests

We declare we have no competing interests.

Funding

This work was supported by the National Institute of General Medical Sciences (NIGMS) of the National Institutes of Health (NIH) grant no. R35GM119856 to K.E.L and NIGMS grant no. R35GM124827 to M.A.W. T.N.P. was supported by NIH-NCI National Cancer Institute grant no. T32CA201160.