Chromosome scale genome assembly for X. cortezi

We generated a nearly chromosome-scale de novo assembly for X. cortezi using PacBio HiFi long-read sequencing at approximately 100× coverage. The genome was highly contiguous, with a contig N50 of 28,997,520 bp. Following reference-guided scaffolding to previously generated chromosome-level X. birchmanni and X. malinche assemblies (NCBI assembly id: GCA_036418095.1), the final X. cortezi assembly was chromosome-level with a scaffold N50 of 32,220,398 bp, and >99.4% of all sequence contained in the largest 24 scaffolds (corresponding to the 24 Xiphophorus chromosomes). The total assembled sequence length of 723 Mb is similar to other Xiphophorus assemblies and close to the expected length for this species based on previously collected flow cytometry estimates [59]. The X. cortezi genome was also highly complete, with 98.6% of actinopterygii BUSCOs present and in single copy (C:98.6%[S:97.0%,D:1.6%],F:0.4%,M:1.0%,n:3640), and the annotation process recovered a total of 25,032 protein coding genes (see Methods, Text A in S1 File). While the 2 genomes are largely syntenic, we also identified putative structural rearrangements between X. birchmanni and X. cortezi (Text A in S1 File and S1 and S2 Tables).

Genome-wide ancestry in the Chapulhuacanito and Santa Cruz populations

Past work from our group has focused on hybridization between X. birchmanni and X. cortezi in the Santa Cruz river drainage [57]. While we collected samples from multiple sites in the Santa Cruz drainage in our previous work, our analyses suggested that hybrids at different sampling sites originated from the same hybridization event [57]. For simplicity, throughout the manuscript, we refer to samples collected at the Santa Cruz site as “Santa Cruz” and samples collected nearby (e.g., in historical collections) as samples from the “Río Santa Cruz.” Here, we report a previously undescribed hybridization event between X. birchmanni and X. cortezi at the Chapulhuacanito population (21°12′10.58″N, 98°40′28.27″W) in the Río San Pedro drainage, 17 km away by land and 130 km away in river distance from the Santa Cruz population (Fig 1A). While on average populations from the Santa Cruz drainage derive 85% to 89% of their genomes from the X. cortezi parental species, the Chapulhuacanito population is more admixed, with 76% of the genome derived from X. cortezi on average (Fig 1A). In both populations, X. birchmanni is the minor parent species.

We also sequenced historical samples from the Chapulhuacanito and Río Santa Cruz populations from 2003, 2006, and 2017. This sampling period spans approximately 40 generations based on reported generation times for this species group [60]. Since hybridization began in these populations more than a hundred generations before the present (see below), our earliest sampling points only survey the latest chapter in the history of X. birchmanni × X. cortezi hybrid populations. Theory predicts that in the first several generations following hybridization, admixture proportions can change dramatically due to selection [20,34,37], but after this initial period, change in genome-wide average ancestry is expected to slow dramatically [37,38]. The observed patterns in our data sets are concordant with these predictions. Genome-wide average ancestry was essentially unchanged from 2003 to recent sampling from 2019 to 2021 (Chapulhuacanito: 78 ± 1.2% X. cortezi in 2003 and 76 ± 2% X. cortezi in 2021; Río Santa Cruz: 87 ± 5% X. cortezi in 2003 and 88 ± 1% X. cortezi in 2019 to 2020).

Demographic history of the hybrid populations

The demographic history of each hybrid population is also expected to impact how repeatable the outcomes of selection are and should be explicitly incorporated into analyses. When comparing repeatability in local ancestry across hybrid populations that formed on different timescales, both the amount of time for selection to shift ancestry at target loci and for genetic drift to shift ancestry at neutral loci would be expected to impact results. To incorporate these sources of variation into our analyses, we used an approximate Bayesian computation approach to explore the likely demographic histories of both the Santa Cruz and Chapulhuacanito populations (see Methods; [23]). We performed simulations drawing from uniform distributions of time since admixture, admixture proportion, and hybrid population size and log uniform distributions for migration rates from each parental species, using SLiM ([61], see Methods). We used the observed genome-wide admixture proportion, coefficient of variance in genome wide and local ancestry, and median ancestry tract length as summary statistics, and used ABCreg ([62]; see Methods) to infer posterior distributions for the time since admixture, admixture proportion, hybrid population size, and migration rates from each parental species, in both hybrid populations. While we did not recover well-resolved posterior distributions for hybrid population size for either population, we do recover well-resolved posterior distributions for other demographic parameters. Based on the maximum a posteriori (or MAP) estimate of these distributions, we find that hybridization began over a hundred generations ago in both drainages (Figs 1A and S1; Chapulhuacanito = 137; Santa Cruz = 263; see S3 Table for 95% confidence intervals), and that the migration rate from the parental populations has been very low (MAP estimate for Santa Cruz: m_cortezi = 4 × 10⁻⁵, m_birchmanni = 0.00028; Chapulhuacanito: m_cortezi = 4 × 10⁻⁵, m_birchmanni = 0.00017; S1 Fig; see S3 Table for 95% confidence intervals). In all subsequent simulations described in the manuscript, we explicitly incorporate this inferred demographic history to build our expectations of cross-population correlations under neutrality or under different models of selection.

Confirming the independent origin of the 2 hybrid populations

Given substantial river distance between the 2 X. birchmanni × X. cortezi hybrid populations (approximately 130 km; Fig 1A) and the limited dispersal capabilities of these small fish ([63,64]; Text B in S1 File), we had good reason to believe that the 2 populations originated independently. However, because of the extraordinarily high correlations in local ancestry we observed across the 2 populations (see below), we sought additional evidence that they were completely independent in origin.

Since we inferred local ancestry for individuals from both populations, we have access to information about historical recombination events in these populations. Specifically, a subset of recombination events that occurred in the hybrid ancestors of present-day individuals will be detectable as ancestry transitions along the chromosomes of present-day individuals. We tested for potential overlap in the locations of ancestry transitions across populations, which could indicate a period of shared history following hybridization. We generated a matrix containing the locations of ancestry transitions in each hybrid individual in our dataset (see Methods) and performed a principal component analysis. We see that the Santa Cruz and Chapulhuacanito populations separate out in PC space in this analysis (Fig 1C). This suggests that the 2 populations have largely distinct recombination histories in the recent past. We also find that the frequency at which the locations of ancestry transitions are shared between individuals in the Santa Cruz and Chapulhuacanito populations is similar to the frequency expected under a null model (Fig 1D), again pointing to independent population histories.

We further explored these patterns using high-coverage whole genome sequencing data of 3 individuals from sympatric X. birchmanni populations at both Santa Cruz and Chapulhuacanito, and 3 naturally occurring hybrids at the 2 sites. We called variants (see Methods) and performed principal component analysis on sympatric X. birchmanni individuals, hybrid samples, and pure X. birchmanni and X. cortezi collected from allopatric populations (S2–S4 Figs). Moreover, we performed local ancestry inference on the natural hybrids for which we had generated deep-sequencing data and identified homozygous X. birchmanni and homozygous X. cortezi ancestry tracts within these individuals (limiting our analysis to tracts that were the same ancestry state in all 6 deep-sequenced hybrids). We extracted these regions from the natural hybrids and from the parental genomes and performed principal component analysis on regions of X. birchmanni and X. cortezi ancestry separately. We found that ancestry tracts derived from the 2 hybrid populations formed separate clusters (Fig 1E and 1F) and individuals from the 2 hybrid populations have high levels of sequence mismatch, compared to the levels observed within populations (Fig 1G; Methods). Simulations indicate that the degree of sequence mismatch observed across the 2 hybrid populations (Fig 1G) greatly exceeds what would be expected if these populations separated recently (i.e., in the ~200 generations since initial hybridization; S5 Fig; Text C in S1 File).

We next used direct measures of relatedness to search for evidence of close relatives across different populations, which would be indicative of unexpected connectedness between them. For each ancestry source we estimated a genetic relatedness matrix [65] using only regions homozygous for that ancestry source from high-coverage hybrid individuals. We find no evidence of close relatedness across hybrid populations in these regions (see Methods; S6 Fig). We also used IBDseq [66] to evaluate evidence of shared IBD tracts within and between populations in our high-coverage data (although we note that we have limited power given our sample size; see Methods and Text D in S1 File). While we found strong evidence of IBD tracts [24] within both hybrid populations and within X. birchmanni and X. cortezi populations (S7 and S8 Figs and S4 Table), we found very limited evidence for IBD tracts shared across the 2 hybrid populations (S7 Fig). Together, these results corroborate our expectations from geographic isolation and demographic analyses, indicating that hybrid populations in the Santa Cruz and Chapulhuacanito rivers have originated and evolved independently. See Text D and E in S1 File for a more thorough discussion of the implications of analyses of relatedness and genetic variation within and between populations.

Correlations between minor parent ancestry, the local recombination rate, and the density of coding and conserved basepairs

Past work on hybrid populations of Xiphophorus and in other systems [3,19,30,32,35,57] has found that the frequency of minor parent ancestry in the genome often correlates with factors such as the local recombination rate and the density of functional basepairs (e.g., coding regions). In the presence of selection against minor parent ancestry (due to hybrid incompatibilities or other mechanisms; [20,34]), both theory and simulations [37] predict that the level of minor parent ancestry will be positively correlated with the local recombination rate. Similarly, if selected sites fall more frequently in coding (or conserved) regions of the genome, and selection is sufficiently polygenic, we might expect to see a depletion of minor parent ancestry in these regions.

We tested for correlations between the local recombination rate estimated in X. birchmanni and local ancestry along the genome in a range of window sizes in both Chapulhuacanito and Santa Cruz (Fig 2A and S5 Table). Although we have developed recombination maps for both species (see Methods), we chose to use the X. birchmanni map because it is likely to be more accurate (see [35]; Text F in S1 File) and our analyses suggest that it is extremely similar to the X. cortezi map (S9 and S10 Figs and Text F in S1 File). Regardless of window size, we observe strong positive correlations between the local recombination rate and average minor parent ancestry in both populations (Fig 2A and S5 Table). After controlling for the strong effects of local recombination rate, we find that the density of coding (and conserved) basepairs also correlates with the distribution of minor parent ancestry in Chapulhuacanito and Santa Cruz (S6 and S7 Tables; see also [57]). In particular, regions of the genome with an especially high density of coding (or conserved) basepairs appear to be depleted in minor parent ancestry (Fig 2B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

(A) Minor parent ancestry in the Chapulhuacanito population is strongly correlated with the local recombination rate. Here, ancestry and recombination are summarized in 250 kb windows (see also Fig 3C for wavelet-based analysis). (B) After accounting for the strong effect of recombination rate by summarizing ancestry in 0.25 cM windows, we also find that minor parent ancestry is depleted in regions of the genome linked to large numbers of coding or conserved basepairs. We previously reported similar results for the Santa Cruz population for both recombination rate and functional basepair density [57] and for X. birchmanni × X. malinche hybrid populations [35]. (C) Average minor parent ancestry is strikingly correlated across the Santa Cruz and Chapulhuacanito populations. Shown here are analyses of 0.5 cM windows (Spearman’s ρ = 0.82, p < 10⁻¹⁰⁰); these results are observed across all spatial scales tested in both physical and genetic distance (see S12 Fig and S8 and S9 Tables). (D) By contrast, minor parent ancestry is substantially less correlated between 2 X. birchmanni × X. malinche hybrid populations. Shown here are analyses of 0.5 cM windows (Spearman’s ρ = 0.31, p < 10⁻²⁰). For additional comparisons of ancestry in X. birchmanni × X. malinche hybrid populations, see [35,57]. The data underlying this figure can be found in Dryad repository doi:10.5061/dryad.qnk98sfq1.

https://doi.org/10.1371/journal.pbio.3002742.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3.

(A) Example of large shared minor parent ancestry deserts identified on chromosome 22 (tan) as well as shared minor parent ancestry islands (peach) in Chapulhuacanito and Santa Cruz. Note also large regions of low minor parent ancestry at approximately 25–30 Mb found across both populations that do not pass the threshold for being designated as shared ancestry deserts (light gray; in this case the region exceeds the 5% threshold for Santa Cruz). Dashed lines indicate the average ancestry genome-wide and dotted lines represent lower and upper 10% quantiles of minor parent ancestry. (B) Spatial wavelet decomposition of the overall Pearson correlation between inferred minor parent ancestry in Chapulhuacanito vs. Santa Cruz (CHPL vs. STAC) measured at a resolution of 32 kb. The contribution of a given spatial scale is a weighted correlation of wavelet coefficients for the 2 signals at that scale, weighted by the portion of the total variance attributable to that scale (see Methods). Correlations among chromosome means also contribute (chrom), as well as a leftover component (scl) due to irregularity of chromosome lengths. (C) Wavelet correlations between inferred minor parent ancestry and recombination rate for both Chapulhuacanito and Santa Cruz populations. Note that here correlations at each scale are not weighted by variances at the corresponding scales. Points are weighted averages across chromosomes with error bars representing 95% jackknife confidence intervals. (D) Wavelet correlations between inferred minor parent ancestry proportion in cross population comparisons between hybrids derived from the same hybridizing pair (CHPL vs. STAC—X. birchmanni × X. cortezi; ACUA vs. AGCZ–X. birchmanni × X. malinche) and from different hybridizing pairs (CHPL—X. birchmanni × X. cortezi vs. ACUA—X. birchmanni × X. malinche hybrids at the Acuapa site). For visualization, we omit the confidence interval for the wavelet correlation of ancestry in the 2 X. birchmanni × X. malinche populations (ACUA vs. AGZC) at the largest scale, since it spans nearly from −1 to 1. Note that the identity of the minor parent species differs across hybrid population types (X. birchmanni in Chapulhuacanito and Santa Cruz and X. malinche in Acuapa and Aguazarca). The data underlying this figure can be found in Dryad repository doi:10.5061/dryad.qnk98sfq1.

https://doi.org/10.1371/journal.pbio.3002742.g003

Repeatability in local ancestry between replicate hybrid populations

We found that local ancestry along the genome was surprisingly repeatable across the 2 X. birchmanni × X. cortezi hybrid populations (Fig 2C). That is, the observed minor parent ancestry in a given 100 kb region of the genome in one population was highly predictive of the observed minor parent ancestry in that same region in the other population (Spearman’s ρ = 0.79; p = 2 × 10⁻¹⁷¹). We note that because adjacent windows are not independent, for all analyses we report p-values after thinning data to include only 1 window per Mb (admixture LD in both populations decays to background levels over this distance; S11 Fig). The observed correlations in ancestry across populations exceed what we have previously detected in replicate X. birchmanni × X. malinche hybrid populations (Fig 2D). While we detected these patterns across window sizes, they generally increased with larger window sizes (S8 Table). We find that these correlations are robust to controlling for shared features of genome architecture like the local recombination rate and the locations of coding and conserved basepairs using a partial correlation approach (S9 Table; see Methods).

Simulations incorporating the inferred demographic history of the 2 populations indicate that this strong cross-predictability is not expected under neutrality but can be produced in scenarios of hybridization followed by strong selection on many loci (Text G in S1 File). This suggests that repeatability in minor parent ancestry across X. birchmanni × X. cortezi hybrid populations is driven by a shared architecture of selection on hybrids. For comparison, we evaluated correlations in local ancestry observed when subsampling individuals from the same population and sampling year, samples from the same populations but different sampling years, and populations sampled from different sites on the same river. We reasoned that for each of these comparisons, samples are expected to largely share the same demographic history and history of selection. Reassuringly, we found that correlations in these analyses greatly exceeded those observed between Chapulhuacanito and Santa Cruz (S12 Fig and S8 Table).

Our simulations indicate that the striking correlations we see in local ancestry across Chapulhuacanito and Santa Cruz (Fig 2C) could be driven by a shared architecture of selection on hybrids in these populations (see below, Text G in S1 File). However, we wanted to thoroughly rule out other possible explanations, namely that technical factors might contribute to this signal. These approaches are described in detail in the Methods and S1 File, but we discuss them briefly here. We used simulations and analyses of lab generated crosses to confirm that our local ancestry inference approach is highly accurate (S13–S15 Figs; Methods and Text H in S1 File). We used simulations to artificially induce high error rates in local ancestry inference and found that it could not generate the patterns observed in our data (Text H in S1 File). We repeated analyses removing regions that are prone to error in local ancestry inference (S10 Table; Text I in S1 File) and controlling for our power to infer local ancestry along the genome (see Methods; S10 Table), among other analyses (see S10 Table; Methods; Text I in S1 File). None of these analyses qualitatively changed our results (Text G–I in S1 File).

Simulations indicate that it is possible for selection alone to drive cross-population correlations at the magnitude we infer in Santa Cruz and Chapulhuacanito in scenarios where selection acts on many loci and is exceptionally strong (average s drawn from an exponential distribution of 0.4–0.6; Text G in S1 File). Our results from lab-generated X. birchmanni × X. cortezi hybrids indicate that hybrids in this cross suffer immense fitness consequences, suggesting that such strong selection is plausible (see Discussion; [67]). Indeed, evaluating patterns of local ancestry across the 2 independently formed populations, we can see evidence for large, shared deserts of minor parent ancestry (Fig 3A). This hints that the correlations we observe in our data may be largely driven by strong selection acting shortly after hybridization, resulting in shared patterning of minor ancestry over broad spatial scales along the genome. To evaluate this question in more depth and across spatial scales in the genome, we next used a wavelet-based analysis of cross population ancestry correlations [38].

Wavelet transform approach to infer the spatial scale of correlations in ancestry

In our windowed analyses, the correlations in ancestry between the Santa Cruz and Chapulhuacanito populations increase as we consider larger window sizes, suggesting that the observed correlations are driven by covariation in ancestry at large genomic scales (S8 Table). Similarly, we find that the correlations between recombination rate and minor parent ancestry become stronger in larger genomic windows (S5 Table).

Theory predicts that the strength of selection on hybrids will vary dramatically over time, since the removal of ancestry tracts harboring alleles that are deleterious in hybrids will be most rapid in the earliest generations following hybridization when ancestry tracts are long [34,36,37]. Furthermore, these dynamics can establish spatial ancestry patterns along the genome that persist over time and constrain subsequent evolution. This leads to the prediction that the genomic scale of autocorrelation in ancestry will be informative about the timing and strength of selection (relative to the onset of hybridization) [38]. To better understand the role of selection in shaping genomic ancestry patterns across replicate hybrid populations, we applied recently developed methods based on the Discrete Wavelet Transform [38] to our data (see Methods; Text J in S1 File). The intuition behind this analysis is as follows: moving along a chromosome, the ancestry proportion deviates around its chromosome-wide average, and this variation occurs over a range of different spatial scales (genomic window sizes, roughly speaking). The wavelet transform can be used to summarize the scales of variance in ancestry along a chromosome, as well as the contributions of each scale to the overall correlation between 2 signals measured along the chromosome (e.g., ancestry and recombination), where each component carries independent information about the overall correlation (see Methods). Because the scale of variation is ultimately determined by the lengths of admixture tracts, these signals contain information about the timing of selection and drift relative to the onset of hybridization [38].

Using this approach, we found that the overall correlation between minor parent ancestry and recombination in both replicate populations is predominantly attributable to broad genomic scales (Fig 3C). Furthermore, wavelet correlations between minor parent ancestry and recombination were strongly positive in replicate populations, with the strongest correlations observed at the broadest genomic scales (S16 Fig). The squared correlation coefficients for ancestry versus recombination can be interpreted as the percent of variance in ancestry at each scale attributable to genome-wide selection against minor parent ancestry [38], since these correlations are only generated by selection and not by drift (barring errors in ancestry inference; see Text J in S1 File). Applying this logic, correlations with recombination indicate that roughly 80% of the variance in ancestry at the broadest genomic scales (e.g., 16 Mb) in the Santa Cruz and Chapulhuacanito populations can be attributed to selection against minor parent ancestry. By contrast, comparatively little of the variance in ancestry at fine genomic scales is attributable to selection against minor parent ancestry (e.g., 0.2% at a scale of 32 kb).

We next applied this approach to investigate the correlation between minor parent ancestry across the 2 replicate X. birchmanni and X. cortezi populations. We found that across scales, cross-population ancestry correlations between X. birchmanni × X. cortezi hybrid populations were stronger than the correlations observed with recombination rate, especially at finer spatial scales. Thus, ancestry in a replicate hybrid population is a better predictor of fine-scale genetic ancestry patterns than recombination rate. This implies that recombination alone only captures a portion of the total effects of selection on ancestry patterns, and that its effects in mediating parallel genomic outcomes of hybridization manifest predominantly over broad genomic scales. From cross-population ancestry wavelet correlations, estimates of the proportion of ancestry variance attributable to selection on minor parent ancestry range from approximately 25% at a scale of 32 kb to as high as 93% at a scale of 8 Mb (Fig 3D). Surprisingly, we found that significant positive correlations persisted even at very small spatial scales (S17 Fig). This pattern is consistent with convergent selection shaping very fine scale ancestry patterns, although we discuss important caveats to this interpretation in Text J in S1 File. Nonetheless, the magnitude and scales of ancestry correlations across populations suggest that predictability is driven by both early and continued selection on hybrids.

For comparison, we repeated these analyses in 2 X. birchmanni × X. malinche hybrid populations, the Acuapa population and the Aguazarca population [35]. X. birchmanni and X. malinche are more closely related than X. birchmanni and X. cortezi (Fig 1B), and hybridization began more recently (within the last 50 to 100 generations; [23]). We again find strong positive correlations between minor parent ancestry in the 2 populations at broad genomic scales, but these are noticeably reduced relative to the cross-population comparison between the 2 X. birchmanni × X. cortezi populations (Figs 3D and S17 and Text J in S1 File). These results are consistent with weaker selection overall against minor parent ancestry in X. birchmanni × X. malinche hybrid populations, and/or fewer loci under selection, both of which may be expected given that these species diverged more recently (Fig 1B; [35,68]). Moreover, previous work analyzing wavelet correlations between minor parent ancestry and recombination rate in X. birchmanni × X. malinche populations found that only ~20% of the variation in minor parent ancestry at large spatial scales was attributable to selection [38]. Overall, these results suggest that genome evolution after hybridization is substantially more predictable for X. birchmanni × X. cortezi hybrids.

Finally, we examined the genomic scale of shared ancestry patterns between a X. birchmanni × X. cortezi hybrid population and a X. birchmanni × X. malinche hybrid population (the Chapulhuacanito and Acuapa populations, respectively). We observed positive correlations in minor parent ancestry at broad scales but find that these correlations are dramatically reduced at fine scales, especially compared to analyses of the populations of the same hybridizing pair (Fig 3D). This would be expected if replicate populations of the same hybridizing pair show greater overlap in the fine-scale targets of selection than populations of different hybridizing pairs. The positive fine-scale ancestry correlations within replicate hybrid populations (Figs 3D and S17) are consistent with this interpretation (see also Text J in S1 File). We thus suggest that broad-scale predictability among different hybridizing pairs may be driven primarily by effects of shared genome architecture rather than shared identity of selected loci.

Repeatability in minor parent deserts and islands between replicate X. birchmanni × X. cortezi populations

Given that the results of wavelet-based analyses point to shared targets of selection across X. birchmanni × X. cortezi hybrid populations, we were interested in whether we could identify individual loci that are likely to be under selection. Loci that are shared targets of selection could be alleles that are globally deleterious (or beneficial), or those that are involved in hybrid incompatibilities between X. birchmanni × X. cortezi. Using our large recent population samples, we identified contiguous regions of low minor parent ancestry, or minor parent ancestry “deserts,” in each X. birchmanni × X. cortezi hybrid population and asked how frequently they overlapped across populations (see Methods). Simulations suggest that our approach has high sensitivity and low false positive rates (approximately 70% power at s = 0.05; average of 2 to 4 shared deserts detected genome wide in neutral regions; see Text K in S1 File). We identified 115 “deserts” of low minor parent ancestry in Santa Cruz and 152 deserts in Chapulhuacanito. Strikingly, 38 of these regions overlapped across the 2 populations, greatly exceeding expectations by chance (Fig 4A; p < 0.001 by simulation; see Methods). The average length of these regions was 1.8 Mb with a total of approximately 40 Mb of the 723 Mb genome falling into shared deserts. Since the typical ancestry tract length for X. cortezi (i.e., the major parent) in these populations is much smaller (approximately 150 kb), this hints that these regions may have changed in ancestry shortly after initial hybridization. These shared minor parent ancestry deserts are excellent candidates for shared regions under selection in the 2 hybrid populations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4.

(A) Shared minor parent ancestry deserts and islands in X. birchmanni × X. cortezi populations (Chapulhuacanito and Santa Cruz—colored points) occur at a much higher rate than expected by chance (gray points, see methods). Black diamonds show the observed number of shared minor parent deserts or islands across the 2 populations and colored points show the results of jack-knife bootstrapping the data from each population in 10 cM blocks (circles–bootstrap results from Chapulhuacanito, triangles–bootstrap results from Santa Cruz). (B) Ancestry of shared minor parent deserts (left) and islands (right) through time in the Chapulhuacanito data set. Points show ancestry at individual deserts or islands for each sampling year, and lines connect results for a given desert or island across years. Shared minor parent ancestry deserts were largely fixed by the onset of genetic monitoring of these populations approximately 40 generations ago. Islands also tended to have high minor parent frequency at the onset of sampling, but several islands do change significantly in minor parent ancestry over the sampling period (S11 Table). Islands that increase significantly in minor parent ancestry through time are outlined in black. (C) One shared minor parent ancestry desert on chromosome 6 overlaps with a known mitonuclear incompatibility generated by combining the X. cortezi mitochondria with homozygous X. birchmanni ancestry at ndufa13 [58,69]. Local ancestry along chromosome 6 in Chapulhuacanito is shown in the top plot and local ancestry along chromosome 6 in Santa Cruz is shown in the bottom plot. The locations of shared deserts and islands are highlighted in tan and peach, respectively. The location of ndufa13 is indicated by the red triangle. Dashed lines indicate the average ancestry genome-wide and dotted lines represent lower and upper 10% quantiles of minor parent ancestry. (D) Both ndufa13 and another gene involved in mitonuclear incompatibility between X. cortezi and X. birchmanni, ndufs5, are nearly fixed for major parent ancestry at the onset of our time-series sampling. (E) With our new long-read reference assemblies, we evaluated minor parent ancestry at the center of inversions (focusing on inversions >100 kb) that differentiated X. birchmanni and X. cortezi in the 2 hybrid populations. Example alignment of a large inversion identified on chromosome 8 is shown in the inset. For each inversion, we sampled ancestry in a 50 kb window that overlapped with the center of the inversion (schematically shown by the gray rectangle in the inset). We found that minor parent ancestry was modestly depleted in the 2 hybrid populations (colored points and whiskers) at inversions compared to randomly sampled regions of the genome (gray points–“all regions”). However, when we generated null data sets only from regions of the genome with low recombination rates (lowest 5% quantile of recombination rate), we found that inversions did not show unusually high depletion of minor parent ancestry. This suggests that depletion of minor parent ancestry at inversions may be driven by reduced recombination in these regions in hybrids. The data underlying this figure can be found in Dryad repository doi:10.5061/dryad.qnk98sfq1.

https://doi.org/10.1371/journal.pbio.3002742.g004

Similarly, we identified regions of especially high minor parent ancestry in each X. birchmanni × X. cortezi hybrid population and asked how frequently they overlapped across populations compared to expectations by chance (see Methods). In doing so, we found evidence for 89 shared minor parent “islands” out of 238 islands in Santa Cruz and 147 in Chapulhuacanito, again exceeding the level of sharing expected by chance (Fig 4A; p < 0.001 by simulation; Methods). The typical length of shared islands was 190 kb, much smaller than that observed for shared deserts, but together these regions still covered a substantial portion of the genome (approximately 29 Mb). We report the genes observed in these regions (S11 Table) and analysis of functional enrichment in the supplementary materials (Text L in S1 File).

We compared minor parent deserts and islands identified in the X. birchmanni × X. cortezi hybrid populations to those detected in the X. birchmanni × X. malinche hybrid populations. As expected, we found many fewer shared deserts and islands across hybrid population types (S18 Fig), but shared deserts and islands exceeded expectations by chance in most comparisons (most comparisons p < 0.001 by simulation, with a maximum of p = 0.11).

Since we had access to time-series data for both the Santa Cruz and Chapulhuacanito populations, we were interested in evaluating how ancestry at minor parent deserts and islands has changed over the last 40 generations. Given that both hybrid populations are estimated to be over 100 generations old, we would expect that loci under strong or moderate selection would be fixed even at the earliest time points in our data set. Indeed, we find that regions that fall into shared ancestry deserts tend to have low minor parent ancestry in 2003 and maintain low ancestry through time (Fig 4B and 4D). The same is generally true for regions of high minor parent ancestry, although we do identify 6 minor parent islands where minor parent ancestry significantly increases between 2003 and 2020–2021 (S11 Table).

Finally, we evaluated ancestry at rearrangements identified between X. birchmanni and X. cortezi based on our new PacBio HiFi-based assemblies. We identified 9 inversions greater than 100 kb, ranging in size from 218 kb to 6.7 Mb (S2 Table and S19 Fig). These inversions were concentrated on chromosomes 8 and 17 (6 out of 9 of the inversions). As chromosomal inversions tend to suppress recombination in heterozygotes, we predicted that these regions would be especially depleted in minor parent ancestry. Notably, we found that on average these regions were depleted in minor parent ancestry compared to expectations from null simulations (Fig 4E), but not when compared to non-inverted regions of the genome that had exceptionally low recombination rates (Fig 4E).

Ancestry at known incompatibilities identified between X. birchmanni and X. cortezi

We were also interested in evaluating patterns of minor parent ancestry locally at regions that are known to be under selection in hybrids between X. birchmanni and X. cortezi. Other work from our lab has identified a mitonuclear hybrid incompatibility between individuals with the X. cortezi mitochondria and homozygous X. birchmanni ancestry at ndufs5 and ndufa13 [57,58,67]. F₂ hybrids that inherit the X. cortezi mitochondrial haplotype and 2 copies of the X. birchmanni allele at ndufs5 experience mortality during embryonic development [58,69]. Inheriting the X. cortezi mitochondrial haplotype and 2 copies of the X. birchmanni allele at ndufa13 causes higher rates of postnatal mortality. Because hybrid populations at both Santa Cruz and Chapulhuacanito have fixed the X. cortezi mitochondrial haplotype (S12 Table; [57,58]), this leads to the strong expectation that they will largely have purged X. birchmanni ancestry at ndufs5 and ndufa13.

We evaluated ancestry in these regions of the genome in our large sample of hybrid individuals from both Santa Cruz and Chapulhuacanito. We identified a large, shared ancestry desert surrounding ndufa13 on chromosome 6 (Fig 4C). For ndufs5, the region surrounding the gene on chromosome 13 was identified as an ancestry desert in Chapulhuacanito, but not in Santa Cruz. Closer examination of this region (S20 Fig) indicates that X. birchmanni ancestry at ndufs5 is depleted in Santa Cruz but falls just above the 5% quantile of minor parent ancestry used to identify deserts genome-wide in Santa Cruz (the 5% quantile was 2.2% X. birchmanni ancestry while an average of 2.3% X. birchmanni ancestry was observed at ndufs5; see Methods). Moreover, both regions were consistently low in X. birchmanni ancestry through time in our samples from Chapulhuacanito (Fig 4D) and no individuals homozygous for X. birchmanni ancestry at either region were observed across the 2 populations. Based on predictions from Hardy–Weinberg equilibrium <0.05% of mating events would be expected to produce embryos incompatible at ndufs5 or ndufa13 in either population. Since these 2 genes form part of mitochondrial protein complex I, we also analyzed ancestry at genes that are involved in protein complexes genome wide (S21 Fig and Text M in S1 File).

Results from a recent study identified 2 additional regions on chromosomes 7 and 14 that impact viability in F₂ hybrids between X. birchmanni and X. cortezi [67]. While these regions are currently poorly resolved given the size of the mapping panel and the use of early generation hybrids (where ancestry tracts are large; [67]), we were curious if they overlapped with any minor parent ancestry deserts in Santa Cruz or Chapulhuacanito. We found that the region on chromosome 14 overlaps with a large shared ancestry desert and the region on chromosome 7 overlaps with an ancestry desert found in Santa Cruz (S22 Fig). We note, however, that these findings are not necessarily surprising given the size of the mapped regions that impact viability in F₂ hybrids [67]. For example, the region identified on chromosome 14 in F₂ hybrids is approximately 6 Mb and has a 27% probability of overlapping with a shared desert by chance based on simulations.

Sample collection

Samples for low-coverage whole genome sequencing were collected from 2 different geographical regions (Fig 1). Wild fish were collected using baited minnow traps in Hidalgo and San Luis Potosí, Mexico. We previously identified hybrids between X. birchmanni and X. cortezi at multiple sites on the Río Santa Cruz in northern Hidalgo [23,68]. We continued to sample from these sites for the present analysis (Huextetitla—21°9′43.82″N 98°33′27.19″W and Santa Cruz—21°9′27.63″N 98°31′13.79″W). We also added a new site in a different drainage (Fig 1), near the town of Chapulhuacanito (21°12′10.58″N 98°40′28.27″W). This site also contained X. birchmanni × X. cortezi hybrids (see Results), but this hybridization event is clearly independent given the geographical distance between these locations and other evidence (Text B–D in S1 File). At both collection sites, nearly pure X. birchmanni individuals were also sampled. These individuals were identified based on their genome-wide ancestry and excluded from further analysis.

We combined previously collected data sets from the Río Santa Cruz (N = 254; [23,68]) with 216 new samples collected from Chapulhuacanito in June of 2021. Collected fish were anesthetized in 100 mg/ml buffered MS-222 and water, following Stanford APLAC protocol #33071. A small fin clip was taken from the caudal fin of each individual and preserved in 95% ethanol for later DNA extraction.

For this study, we also took advantage of historical collections from 2003, 2006, and 2017 in the same regions. These samples were matched to present-day collection sites using GPS coordinates and represented a mix of fin clips preserved in DMSO and whole fish preserved in 95% ethanol. We prepared libraries, sequenced all samples, and identified 76 hybrids from historical samples from Chapulhuacanito and 23 from the Río Santa Cruz.

Chromosome scale assembly for X. cortezi

We generated a new reference genome for X. cortezi for this project from a lab-raised male descended from an allopatric population sampled on the Río Huichihuayán. Previous work involving X. cortezi used a draft genome assembled with 10× chromium linked read technology [23,68]. We assembled the new reference using PacBio HiFi data.

Genomic DNA was isolated from tissue using QIAGEN’s Genomic-Tip 500/G columns following the manufacturer’s recommendations with some adaptations. Approximately 400 mg of body tissue was digested in 1.5 ml of Proteinase K and 19 ml Buffer G2 at 50°C for 2 h, inverting the sample every half hour. Following the incubation, the column was equilibrated using 10 ml of Buffer QBT. The sample was vortexed for 10 s at maximum speed, then immediately applied to the column. Two washes were performed with a total of 30 ml of Buffer QC. The column was then transferred to a clean 50 ml tube and genomic DNA was eluted from the column with 15 ml of Buffer QF that was prewarmed to 50°C. The DNA was precipitated using 10.5 ml of isopropanol, mixed gently, then centrifuged immediately at a speed of 5,000 × g for 15 min at 4°C. The DNA pellet was then washed with 4 ml of cold 70% ethanol and re-pelleted via centrifugation. Then, the pellet was air-dried for 10 min and resuspended in 1.5 m of Buffer EB. Genomic DNA was quantified and assessed for quality using a Qubit fluorometer, Nanodrop, and Agilent 4150 TapeStation. Extracted DNA was sent to Admera Health Services, South Plainfield, NJ for PacBio library prep and sequencing on SMRT cells. Raw sequence data is available on NCBI’s Sequence Read Archive (BioProject PRJNA1106506).

To remove residual adapter contamination from the HiFi reads, we used HiFiAdapterFilt [73] with the default match parameter of 97% and a length parameter of 30 bp. We then generated a phased genome assembly with hifiasm (v0.16.1; [74]). The resulting primary assembly was 144 contigs with a N50 of 28,997,520 bp. To achieve a chromosome-level assembly, we scaffolded the X. cortezi genome to the chromosome-level genomes of species in its sister clade: X. birchmanni and X. malinche (NCBI assembly id: GCA_036418095.1) using RagTag (v2.1.0; [75]). Where these scaffolded genomes differed in synteny, we used the chromosome-level assemblies of X. hellerii, X. maculatus, and X. couchianus as outgroups to select the ancestral orientation for X. cortezi. This scaffolded X. cortezi genome had a scaffold N50 of 32,220,398 bp and length of 723,632,656 bp. These putative X. cortezi chromosomes were aligned to the X. maculatus genome assembly using minimap2 (v2.24; [76]) and oriented and numbered according to identity with X. maculatus.

Chromosome 21 is known to contain the major sex determination locus in many Xiphophorus species [77]. To resolve potential structural variation at this locus and include both X and Y linked sequence in the X. cortezi reference genome, we generated an alignment between the 2 inferred haplotypes for chromosome 21. We found that one chromosome 21 haplotype was syntenic to chromosome 21 in X. birchmanni, while the other contained a 7 Mb chromosomal inversion relative to X. birchmanni, which is syntenic to all other Xiphophorus species and likely represents the ancestral Xiphophorus arrangement of the Y-chromosome [78].

The mitochondrial genome was assembled from the adapter-filtered HiFi reads using MitoHiFi (v3.2; [79]) with default parameters and using the X. maculatus mitochondrial genome as a reference. We used BLASTn [80] searches to identify and subsequently remove mitochondrial contaminant sequences present in the nuclear genome, which were present on only 6 contigs that were all less than 40 kb in length. Following contaminant removal, the mitochondrial genome assembled with MitoHiFi was added to the X. cortezi assembly. The final assembly is available on NCBI (Bioproject PRJNA1106506).

Annotation of the X. cortezi assembly

The X. cortezi genome was annotated using a pipeline adapted from a previous study [81]. Transposable elements (TEs) in the assembly were identified using RepeatModeler and RepeatMasker [82]. RepeatModeler was first used for an automated genomic discovery of TE families in the assembly. This result, together with Repbase and FishTEDB [83,84], was input into RepeatMasker for an additional retrieval of TEs based on sequence similarity. For protein coding gene annotation, TEs from known families were hard-masked and simple repeats were soft-masked from the assembly. We used a tool designed to parse RepeatMasker output files [85] to compute quantitative information on representation of different TE families. We repeated this approach for the X. birchmanni PacBio reference assembly generated using the same approach. Analysis of differences between the 2 species in repeat content is available in Text A in S1 File.

Protein coding genes were annotated by collecting and synthesizing gene evidence from homologous alignment, transcriptome mapping, and ab initio prediction. For homologous alignment, 455,817 protein sequences were collected from the vertebrate database of Swiss-Prot (https://www.uniprot.org/statistics/Swiss-Prot), RefSeq database (proteins with ID starting with “NP” from “vertebrate_other”), and the NCBI genome annotation of human (GCF_000001405.39_GRCh38), zebrafish (GCF_000002035.6), platyfish (GCF_002775205.1), medaka (GCF_002234675.1), mummichog (GCF_011125445.2), turquoise killifish (GCF_001465895.1), and guppy (GCF_000633615.1). We then aligned those protein sequences onto the assembly using both GeneWise and Exonerate (https://www.ebi.ac.uk/about/vertebrate-genomics/software/exonerate) to collect homologous gene models. In order to speed up GeneWise, GenblastA was used to retrieve the rough alignment region of the assembly for each protein [86].

For transcriptome mapping, we used previously collected RNA-seq reads from multiple tissues [17], cleaned them using fastp [87], and mapped them to the assembly using HISAT [88]. StringTie was then used to interpret gene models from the mapping results [88]. In parallel, we used Trinity to assemble RNA-seq reads into transcript sequences and aligned them to assembly for gene modeling using Splign [89,90].

We used AUGUSTUS for the ab intio gene prediction [91]. AUGUSTUS was trained for the first round using BUSCO genes. Genes that were predicted repeatedly by Exonerate, Genewise, StringTie, and Splign were considered to be high-quality genes and were used to train AUGUSTUS for the second round. All collected homologous and transcriptome gene evidence were used as hints for AUGUSTUS for the ab initio gene prediction.

To generate the final consensus annotation, we screened homology gene models locus by locus. When 2 gene models competed for a splice side, we kept the one better supported by transcriptome evidence (using transcriptome data from [17]). When a terminal exon (with a start/stop codon) from an ab initio or homology gene model was better supported by transcriptome data than that of the previously selected gene model, the exons in question were replaced by the predictions of the gene model best supported in the transcriptome data. We also kept an ab initio prediction when its transcriptome support was 100% and it had no homology prediction competing for splice sites.

Low coverage whole genome sequencing

We extracted DNA from fin clips collected from wild-caught fish using the Agencourt DNAdvance kit (Beckman Coulter, Brea, California, United States of America). We used half-reactions but otherwise followed the manufacturer’s instructions for DNA extraction. We used a BioTek Synergy H1 (Agilent, Santa Clara, California, USA) microplate reader to quantify extracted DNA. We diluted DNA to a concentration of 10 ng/μl and then prepared tagmentation-based libraries from this genomic DNA for low-coverage whole genome sequencing. The approach used for generating libraries is described in Langdon and colleagues [23]. Dual-indexed libraries were bead purified with 18% SPRI magnetic beads, quantified on a qubit fluorometer (Thermo Scientific, Wilmington, Delaware, USA), and visualized on an Agilent 4200 Tapestation (Agilent, Santa Clara, California, USA). Purified libraries were sequenced by Admera Health Services (South Plainfield, New Jersey, USA) on an Illumina HiSeq 4000 instrument.

Whole genome resequencing

To evaluate patterns of genetic variation within ancestry tracts, we sequenced a subset of individuals (N = 3 per genotype per population) at high coverage. For these individuals, we prepared libraries following the approach of Quail and colleagues [92]. We used 500 ng to 1 μg of DNA per sample and sheared this input DNA to approximately 400 bp using a QSonica sonicator. The fragmented DNA underwent an end-repair reaction with dNTPs, T4 DNA polymerase, Klenow DNA polymerase, and T4 PNK for 30 min at room temperature. An A-tail was added to the end-repaired DNA using a mix of Klenow exonuclease and dATP, incubated for 30 min at 37°C. The A-tail facilitated ligation of adapters with DNA ligase in a 15-min reaction performed at room temperature. The resulting sample was purified using the Qiagen QIAquick PCR purification kit. Barcodes were added during a final PCR amplification step using the Phusion PCR kit, which was run for 12 cycles. This reaction was purified with 18% SPRI beads and libraries were visualized on the Agilent 4200 Tapestation and quantified using a Qubit fluorometer. These libraries were also sent to Admera Health Services for sequencing on an Illumina HiSeq 4000 machine.

Inferring recombination maps for X. birchmanni and X. cortezi

In past work, we used population genetic methods to infer a linkage disequilibrium (LD)-based recombination map for an earlier version of the genome assembly for X. birchmanni [35]. We repeated the same approaches with the new X. birchmanni reference genome to generate a new LD-based map. Briefly, we used the previously published resequencing data for 22 adult X. birchmanni individuals and a pedigreed family with 5 offspring [35], for a total of 24 unrelated adults. We mapped reads to the genome with bwa mem, realigned indels with PicardTools, and called variants with GATK (v3.4; [93]). We filtered variant and invariant sites based on quality thresholds as we had with the original recombination map (DP < 10; RGQ < 20; QD < 10; MQ < 40; FS > 10; SOR > 4; ReadPosRankSum< -8; MQRankSum < -12.5). We excluded sites that overlapped with annotated repetitive regions or had <0.5× or >2× the average genome-wide coverage for that individual. For invariant sites, only RGQ and DP filters could be used. Using this filtered list of sites, we inferred the expected error rate with plink [94] using expectations of mendelian segregation in the pedigree. Finding evidence of a low error rate (approximately 0.45% per SNP across 5 offspring), we first removed these errors and then proceeded to phasing and inferring the LD map. We performed phasing using the program shapeit2 with the duohmm flag for inclusion of family data [95]. Past simulations matching parameters observed in X. birchmanni have suggested that although phasing likely introduces errors, improvements in map resolution outweigh errors introduced by phasing [35].

We inferred the LD map using LDhelmet. LDhelmet relies on a mutation transition matrix for recombination map inference [96] and also can take advantage of distributions of ancestral alleles when computing likelihoods. To infer ancestral alleles for both purposes, we used phylofit [97]. Previous simulations matching parameters observed in X. birchmanni have suggested that this approach results in accurate inference of ancestral sequences [35]. We used previously collected whole genome sequence data from 11 species of Xiphophorus (S13 Table) to infer the likely ancestral basepair at variable sites as described previously [35] using the prequel command [97]. To run phylofit, we provided the aligned sequences and the inferred species tree for these groups of species [98]. For mutation matrix inference based on phylofit output, we used a threshold of 0.99 to convert posterior probabilities for the ancestral basepair to hard calls.

We then used phased haplotypes from all unrelated X. birchmanni individuals (48 haplotypes in total) and the mutation transition matrix to infer an LD-based recombination map with LDhelmet [96]. The total number of SNPs input into LDhelmet was 2,565,331. We first computed a likelihood lookup table for ρ values using a grid table ranging from 0 to 10 (sampling in intervals of 0.01 from 0–1 and 1 from 1–10). We next inferred recombination rates using LDhelmet’s rjMCMC procedure with a block penalty of 50, a burn-in of 100,000, and ran the Markov chain for 1,000,000 iterations. Past work has suggested that a block penalty of 50 improves accuracy for inference of broad scale recombination rates in Xiphophorus [35]. Following map inference, we excluded SNP intervals with implausible high recombination rates (ρ/bp ≥ 0.4) and summarized recombination rates in windows of physical distance ranging from 5 kb to 5 Mb. We also used the local recombination rate estimates and the inferred lengths of each chromosome in cMs to divide the chromosome into windows of genetic distance for certain analyses (Text N in S1 File).

Because we had access to whole genome resequencing data for 9 unrelated X. cortezi individuals from the Huichihuayán river (near the Nacimiento) from previous work [17,23], we decided to supplement this data to build an LD-based map for this species as well. To generate a comparable sample size for this inference, we sequenced an additional 8 individuals following the whole genome resequencing protocol described above. The average coverage for the X. cortezi individuals was approximately 65×, and the range was 19–113×. We inferred an LD-based map for X. cortezi as described above, except that we lacked access to pedigree data for mendelian error correction and phasing.

With the lower sample size and lack of pedigree data, we expected the X. cortezi map to be less accurate than the X. birchmanni map but used it to test general hypotheses. Swordtails have deleted the N-terminal domain of PRDM9 [99] and have a conserved PRDM9 zinc-finger binding domain across the clade. Past work has indicated that swordtails behave as PRDM9 knock-outs with a higher frequency of recombination events near the TSS, CpG islands, and H3K4me3 marks [35,99]. Using the inferred LD map for X. cortezi, we confirmed that we observe elevated recombination rates close to the TSS and H3K4me3 peaks, similar to patterns observed in X. birchmanni (S10 Fig; Text F in S1 File). We note that the median inferred ρ/bp in X. cortezi was substantially higher than in X. birchmanni (0.0027 versus 0.00076). Based on the results of our analyses of historical population sizes (see Text F in S1 File), we expect to see elevated ρ/bp in X. cortezi since ρ reflects 4Ne*r and we infer that X. cortezi has had approximately 2× the effective population size of X. birchmanni over the past 100k generations. However, ρ/bp may also be impacted by a higher error rate in the X. cortezi recombination map given the lack of access to pedigree data.

Changes to the local ancestry inference pipeline

We previously developed approaches for local ancestry inference for hybrids between X. birchmanni and X. cortezi, but we made several improvements upon previous implementations for this project. First, we used a new chromosome scale assembly for X. cortezi generated by PacBio HiFi technology (see above). To more accurately quantify allele frequencies in the parental species, we sampled additional allopatric populations of X. cortezi, which has been less intensively sampled from a genomic perspective than X. birchmanni, and sequenced F₁ hybrids between the 2 species for error correction. We also identified and corrected an error in the ancestryinfer code (https://github.com/Schumerlab/ancestryinfer) that had resulted in a number of ancestry informative sites being erroneously excluded in previous versions of the pipeline.

Using the new assemblies, we identified candidate ancestry informative sites by aligning resequencing data from a high-coverage X. cortezi individual to the X. birchmanni PacBio assembly (as described previously; [17,57,100]), and identifying all sites that were homozygous for different states in this data. We then treated these sites (2.64 million) as potential ancestry informative sites and evaluated their frequency in allopatric X. cortezi and X. birchmanni populations using 1× whole genome sequence data from 90 individuals of each species from 3 source populations each (X. birchmanni: Coacuilco, Talol, Xaltipa; X. cortezi: Puente de Huichihuayán, Octzén, Calle Texacal). We identified sites that had a 98% or greater frequency difference between the 2 species at our filtered set of ancestry informative sites (1,001,684 sites). We used these sites and their observed frequencies in the parental species as input for our ancestry HMM pipeline (ancestryinfer; [100]).

We next took advantage of our lab-generated F₁ hybrids to further filter these ancestry informative sites. We collected approximately 1× whole genome sequence data for 42 F₁ hybrids we generated between X. birchmanni and X. cortezi and analyzed these individuals using the ancestryinfer pipeline, specifying the X. birchmanni reference as genome 1 and the X. cortezi reference as genome 2. We set the error rate to 0.02 for this initial analysis. After running the pipeline, we converted posterior probabilities for each ancestry state into hard-calls using a posterior probability threshold of 0.9. Because F₁ hybrids should be heterozygous for all ancestry informative sites across the genome, we identified ancestry informative sites that were called with high confidence as homozygous X. birchmanni or homozygous X. cortezi and excluded these sites. This resulted in a final set of 995,825 ancestry informative sites which we used for downstream analyses, or a median of one marker every 240 basepairs across the 24 major chromosomes.

We tested the performance of this approach on 30 X. cortezi individuals we had not used in our initial filtering, 12 X. birchmanni individuals, 13 F₁ hybrids, 26 F₂ hybrids, and 5 BC₁ hybrids (backcrossed to X. cortezi) where we have clear expectations for true ancestry. Based on this analysis, we found that performance of the HMM approach was excellent (S13–S15 Figs).

Local ancestry inference and processing for downstream analysis

Using the ancestry informative sites described above, we next proceeded to analyze hybrid individuals from Chapulhuacanito and the Río Santa Cruz using the ancestryinfer pipeline. We inferred local ancestry for 291 individuals from Chapulhuacanito and 277 individuals from the Río Santa Cruz. Because previous analyses have indicated that ancestryinfer is not sensitive to priors for initial admixture time and admixture proportions [100], we set the prior for the genome-wide admixture proportion to 0.5 and the prior for the number of generations since initial admixture to 50. However, we repeated local ancestry inference for both populations following demographic inference using ABCreg (see next section) using priors inferred from this analysis for initial admixture time and admixture proportion. We found that our results were qualitatively unchanged (S10 Table). For all analyses, we used a uniform recombination prior, set to the median per-basepair recombination rate in Morgans inferred for X. birchmanni.

For a number of downstream analyses, it was useful to convert posterior probabilities for different ancestry states into hard-calls. As we have previously, we used a posterior probability threshold of 0.9 or greater to assign an ancestry informative site to a given ancestry state (e.g., homozygous X. birchmanni, heterozygous for ancestry, or homozygous X. cortezi). Ancestry informative sites with lower than a 0.9 probability for any ancestry state were masked. We also filtered out sites that were covered in fewer than 25% of individuals. This resulted in 994,891 sites across the genome in Santa Cruz and 994,906 sites across the genome in Chapulhuacanito for downstream analysis. All local ancestry results are available on Dryad (DOI: 10.5061/dryad.qnk98sfq1).

Consistent with previous work [57,68], a subset of the individuals we sequenced were nearly pure X. birchmanni (>98% of the genome derived from the X. birchmanni parent species). We identified and excluded these individuals from our data set before examining patterns of local ancestry within the 2 hybrid populations, resulting in a data set of 114 hybrid individuals from Chapulhuacanito and 276 from the Río Santa Cruz populations. We summarized minor parent ancestry across individuals by average ancestry hard-calls in non-overlapping windows of a range of sizes (e.g., 100 kb to 500 kb and 0.1 to 0.5 cM).

Demographic inference in the Chapulhuacanito and Santa Cruz populations

To inform our understanding of patterns of local ancestry along the genome in Chapulhuacanito and Santa Cruz, we wanted to better understand the likely demographic history of these populations. To do so, we used a regression-based Approximate Bayesian Computation or ABC approach with the software ABCreg [62]. We previously applied a similar approach to infer the likely demographic history of the Santa Cruz population [23] but repeat it for both populations here taking advantage of our larger empirical data sets and updated local ancestry inference pipeline. All simulations were performed in SLiM [61].

For each simulation, we drew each population demographic parameter from a uniform or log-uniform prior distribution, performed simulations, and calculated summary statistics for the simulation. We recorded the summary statistics and simulated parameters and compared them to the same statistics calculated from the real data. We modeled one chromosome 25 Mb in length with local recombination rates matching those observed on X. birchmanni chromosome 2. We used both global and local metrics as summary statistics (S14 Table). We used the tree sequence recording functionality of SLiM to determine local ancestry of each individual in the hybrid population [101]. To perform each simulation, we used the following steps:

1. We initialized parental populations and formed a hybrid population between them. We determined the admixture proportion by drawing from a uniform prior distribution for the proportion of the genome derived from parent 1 (0.5–1). Similarly, we determined the hybrid population size for the simulation by drawing from a uniform prior ranging from 2–10,000 individuals.
2. We drew a migration rate from each parental species from a log uniform prior distribution (ranging from m = 0%–3% per generation based on previous results [101]).
3. We drew a time since initial admixture parameter from a uniform distribution ranging from 10 to 400 generations.
4. We performed the simulation for the number of generations drawn in step 3 above, implementing migration from the parental species each generation at the rate drawn in step 2.
5. We randomly sampled 69 and 242 individuals from the population to match the number of hybrid individuals sampled in Chapulhuacanito and Santa Cruz in 2021 and 2020, respectively.
6. We generated summary statistics to compare to summary statistics calculated based on the real data.
7. We repeated this procedure until 150,000 simulations had been generated.
8. We ran the program ABCreg with the tolerance parameter set to 0.005.

To evaluate the validity of our approach, we tested how well this procedure worked to infer parameters for a simulated population with known history. We randomly sampled 100 simulations generated as described above and treated these simulations as if they were the real data. We calculated summary statistics and ran ABCreg as described above (excluding these simulations from the full ABCreg data set). We then calculated the 95% quantile of the posterior distribution for each demographic parameter and asked how well this distribution captured the known parameters for the focal simulation. In general, the 95% quantile of the posterior distributions for each test set overlapped with the true value (S25 Fig). However, performance was poorer when we asked how often the true value fell in the 50% quantile of the posterior distribution produced by ABCreg, indicating that we should view MAP estimates as approximate estimates of the likely demographic history of each population (S25 Fig).

For all subsequent simulations modeling population evolution in Chapulhuacanito and Santa Cruz, we based demographic parameters on the results of these simulations. Text G in S1 File for more information.

Repeatable patterns of minor parent ancestry as a function of genomic architecture

Using the LD-based recombination map described above, we evaluated evidence for a correlation between the local recombination rate in windows and minor parent ancestry in those same windows. As previously reported [57], we found strong correlations between local recombination rate and average minor parent ancestry in the Santa Cruz population, with minor parent ancestry being more common in regions of the genome with the highest local recombination rates. We repeated this analysis for the Chapulhuacanito population and replicated this pattern.

Because recombination events in Xiphophorus species appear to disproportionately localize to functionally dense regions of the genome (e.g., transcriptional start sites, CpG islands, and H3K4me3 peaks; [24,99]), we wanted to control for proximity to some of these elements in our analyses. We calculated the number of coding and conserved basepairs in each window and incorporated this into our analysis. We calculated the Spearman’s partial correlation between recombination rate, minor parent ancestry, and coding (or conserved basepairs) across a range of window sizes (S6 Table). We also repeated this analysis calculating average ancestry and the number of coding (or conserved) basepairs in windows of a particular genetic length (0.1 to 0.5 cM; S7 Table; Text N in S1 File).

Cross-population repeatability in ancestry

The Santa Cruz and Chapulhuacanito populations occur in separate river systems and multiple lines of evidence indicate that they originated from independent hybridization events between X. birchmanni and X. cortezi. We wanted to understand the extent to which local ancestry between these 2 populations was correlated. Presumably, correlations that are observed (barring those due to technical artefacts) should be driven by shared sources of selection, either due to shared loci under selection or shared genomic architecture (e.g., similar recombination maps and locations of coding and conserved basepairs between species).

To evaluate this, we used a Spearman’s correlation test implemented in R. We calculated these correlations in windows of a range of physical sizes (100 kb to 500 kb) and genetic sizes (0.1 to 0.5 cM). We performed each of these calculations thinning the data to retain only a single window every Mb or a single window every 1.5 cM (typically approximately 600 kb in Xiphophorus). This analysis should be conservative since admixture LD decays to background levels over ~500 kb in Santa Cruz and Chapulhuacanito (S11 Fig). We found that the cross-population correlations in local ancestry in these analyses were surprisingly high (S8 Table). Given this observation, we sought to exclude several technical factors that might be artificially inflating this correlation.

Because power to infer ancestry will vary along the genome, we wanted to evaluate whether accounting for this power variation impacted the signal we observed. Certain regions of the genome have a higher density of ancestry informative sites between X. birchmanni and X. cortezi. We determined the median distance between ancestry informative sites (240 bp), and thinned markers such that in regions with higher marker frequency, we retained at most 1 marker per 240 bp, and repeated local ancestry inference with this marker set. We also identified and excluded windows in which we have especially low power to infer ancestry (the number of ancestry informative sites fell in the lower 5% quantile of the genome-wide distribution).

We investigated the impact of removing other regions of the genome where we might expect to have a higher error rate in local ancestry inference. Analysis of our assemblies using seqtk telo [102] suggest that some of our chromosomes include assembled telomeric regions (S19 Fig). Since these regions may be especially challenging to analyze, we recalculated cross-population correlations excluding any region within 1 Mb of the end of a chromosome. We also generated a version of the ancestry informative sites excluding markers that overlapped with repetitive regions and recalculated cross-population correlations. We performed a number of other complementary analyses that are described in detail in Text G–I in S1 File.

Overall, our qualitative results were unchanged in each of the modifications described above and in the series of additional analyses described in Text G–I in S1 File (S10 Table). As a sanity check, we also performed analyses where we generated subpopulations from either Santa Cruz or Chapulhuacanito and asked about the observed correlations in ancestry when individuals truly originate from the same population. We also compared correlations in ancestry between samples from the same population over time, and from different sites in the same river. Reassuringly, all of these comparisons yielded correlations in ancestry that exceeded what we observed between independently formed populations at Santa Cruz and Chapulhuacanito (S8 Table and S12 Fig).

Evidence of independent formation of Santa Cruz and Chapulhuacanito

While the Río Santa Cruz and Río San Pedro are separated by over 130 km of river miles, we wanted to perform additional analyses to confirm that they were independent in origin given the strong correlations in local ancestry that we observe across the 2 populations. To do so, we used a combination of approaches. First, we performed principal component analysis of the locations of observed ancestry transitions in Santa Cruz and Chapulhuacanito. If these populations formed and evolved independently, we would expect that observed ancestry transitions (which reflect recombination events in the ancestors of each hybrid individual) would occur largely in different locations across the 2 populations.

To generate a data set for principal component analysis, we first identified the approximate locations of ancestry transitions for each hybrid individual in our data sets from Santa Cruz and Chapulhuacanito, defined as the interval over which the posterior probability changes from 0.9 posterior probability for one ancestry state to 0.9 for another ancestry state. Ancestry transitions that were supported by flanking segments of <5 kb were removed. Qualitatively, this removed transitions where the ancestry state switched and then immediately reverted, which we hypothesized were likely to be errors. Once ancestry transitions have been identified, we binned the genome into windows of 0.5 cM, and for a given individual recorded a 1 if there was a transition observed in that window and a zero if there was not. While most ancestry transitions were well resolved (mean 13.6 kb), some spanned multiple windows, so we used the midpoint as the location of the ancestry transition for these purposes. We ran a principal component analysis of this matrix in R.

While the variation in ancestry transition locations suggested broadly different histories of recombination in the Santa Cruz and Chapulhuacanito populations, this pattern could also be consistent with an initial period of shared history followed by vicariance. Thus, as a complementary approach, we quantified how frequently the locations of ancestry transitions were shared between pairs of individuals in the Santa Cruz and Chapulhuacanito populations, compared to expectations by chance. Shared ancestry transitions, reflecting ancestral recombination events, could occur by chance due to recombination hotspots or poor resolution of the precise locations of recombination events, but an excess of shared transitions is likely to reflect shared ancestors and thus a shared population history. For both the real and simulated data, we excluded ancestry transitions that were poorly resolved (>250 kb in length) from our analysis. For the real data, we quantified how frequently the intervals of ancestry transitions overlapped in pairs of individuals from the 2 populations using bedtools [103], treating ≥1 basepair shared as an overlap. To generate simulated data, we performed a series of steps. We first excluded windows where ancestry for 1 parental species had fixed in the real data, as these regions of the genome cannot contain ancestry transitions in the real data. For each individual, we iterated through all of the ancestry transitions observed, randomly sampling a new location for each ancestry transition, weighted by the X. birchmanni recombination map (summarized in 100 kb windows). Within that randomly selected window, we used R’s runif function to identify a start position of the recombination interval and set the stop position based on the interval length. We repeated this until all ancestry transitions for an individual had been assigned a random position weighted by the recombination map and repeated this process for all individuals. Next, we quantified the overlap of ancestry transitions in pairs of simulated Santa Cruz and Chapulhuacanito individuals relative to the real data. These results are shown in Fig 1D.

To further exclude the possibility that observed patterns might be generated by an initial period of shared history followed by vicariance, we performed several additional analyses. We collected high-coverage whole genome sequencing data (>20×) for 3 hybrid individuals from each population and 3 pure X. birchmanni individuals found in the same populations (i.e., individuals inferred to derive >98% of their genome from X. birchmanni). For these individuals, we called variants throughout the genome as described above for inference of LD-based recombination maps (mapping and variant calling were performed using both references independently, and results were qualitatively similar). Using this variant information, we performed a principal component analysis on the observed variants genome wide in the individuals collected from Santa Cruz and Chapulhuacanito as well as previously collected high-coverage data from source populations of X. birchmanni (Coacuilco; [35]) and X. cortezi (Huichihuayán and Puente de Huichihuayán; [57,68]).

Because hybrids combine the genomes of the X. birchmanni and X. cortezi individuals that contributed to the hybridization event, we were also interested in subsetting these regions of the genome and analyzing them separately. To do so, we conducted local ancestry inference on the 6 hybrid individuals as described above, and identified regions where we had high confidence that all 6 hybrid individuals in our data set were homozygous X. cortezi or homozygous X. birchmanni. We extracted the variants in these segments from hybrids and from the corresponding parental species plink files (i.e., from X. cortezi individuals for analysis of homozygous X. cortezi ancestry tracts). We performed a separate PCA on the X. cortezi and X. birchmanni-derived regions in hybrids. If Santa Cruz and Chapulhuacanito somehow shared a population history, we would expect these regions to cluster closely together and potentially overlap in a principal component analysis. See Text E in S1 File for a more detailed discussion of these results and their implications.

To more closely evaluate possible relatedness in these ancestry tracts, we used the program GCTA to generate a genetic relatedness matrix for these 6 hybrid individuals [65]. We performed separate analyses for the X. cortezi and X. birchmanni ancestry tracts. As above, we only analyzed regions where all 6 hybrid individuals were homozygous X. cortezi or X. birchmanni in that region, respectively. We included data from the relevant parental populations for comparison. We found that all hybrid individuals from the same population were inferred to have some degree of relatedness based on analysis of both X. cortezi and X. birchmanni derived ancestry tracts, but all values for cross-population comparisons were negative (S6 Fig).

We also used our high-coverage data from the parental and hybrid populations to infer evidence for IBD tracts using the program IBDseq [66]. Because we do not have access to a large number of high-coverage individuals, we were unsure whether programs designed for large data sets like IBDseq would have reasonable performance on our data. We thus first analyzed a known X. birchmanni pedigree [35] with IBDseq using the parameters errormax = 0.0025, r2window = 100, and r2-max = 0.5 (we use a higher than recommended r2-max due to high admixture LD in our focal populations). We used an LOD threshold of 3 as a cutoff for reporting an IBD tract [66]. When applied to the X. birchmanni pedigree data, we found that IBDseq had excellent performance. We thus applied the same approach to identify IBD tracts in the 6 high-coverage individuals from Chapulhuacanito and Santa Cruz, and in different iterations of the analysis also included high-coverage individuals from allopatric and sympatric parental populations (see Text D in S1 File). However, we note that with our small sample size we might expect more variable performance in IBD tract identification in a sample of less closely related individuals.

Since most of the genome of individuals in the 2 hybrid populations is derived from X. cortezi, we performed an additional analysis focusing on X. cortezi ancestry tracts in the high-coverage individuals. Reasoning that populations derived from distinct source populations and with independent demographic histories should harbor distinct frequencies of genetic variants, we performed a “mismatch” analysis. We subset our data to focus only on regions that were homozygous X. cortezi across all 6 high-coverage hybrid individuals. For each pair of individuals in our data set, we counted each site where individual 1 was homozygous for one allele and individual 2 was homozygous for another (within X. cortezi ancestry tracts). We counted the total number of these sites along the genome, divided by the total number of sites that passed quality thresholds in both individuals (within X. cortezi ancestry tracts), and treated this as our mismatch statistic. We compared this mismatch statistic within-populations versus between-populations (Fig 1G) and to population genetic simulations in SLiM modeling distinct population histories (Text C in S1 File).

Spatial scale of cross-population correlations in ancestry

In the absence of selection or genetic drift, the ancestry proportion in a hybrid population would remain uniform along the chromosome. In real populations, ancestry varies along the genome due to the combined effects of recombination with genetic drift, selection, and repeated admixture events. The spatial scale of ancestry variation along the genome holds important information about the timing of demographic and selective events, since recombination progressively shortens ancestry tracts across generations. We took advantage of the recent application of the Discrete Wavelet Transform to decompose correlations between genomic signals into independent components associated with different spatial genomic scales [38]. Briefly, the method transforms a signal measured along the genome (e.g., ancestry) into a set of coefficients that measure changes in the signal between adjacent windows at different locations and with windows of different sizes. The wavelet transform is performed on the 2 signals separately, and the correlation between the coefficients at a given scale for the 2 signals are weighted by the variance at that scale (also determined from the wavelet coefficients) to give the contribution of each scale to the overall correlation. This approach offers the advantage that correlations across scales carry independent information, in contrast to traditional window-based analyses used elsewhere in the manuscript where results across different window sizes are confounded due to the nestedness of windows of different sizes.

As this analysis requires evenly spaced measurements along a chromosome, we first interpolated admixture proportions within diploid individuals to a 1 kb grid for each chromosome, then averaged across individuals to obtain the interpolated sample admixture proportion. We used the inferred recombination maps described above to obtain estimates of recombination in windows centered on the interpolated ancestry measurements. We applied a threshold to recombination values of ρ ≥ 0.005 (corresponding to 4% of the genome) which we found improved the strength of correlation between genetic lengths of chromosome inferred from the LD map versus from an F₂ linkage map.

We used the R package gnomwav [38] to estimate wavelet correlations between signals (minor parent ancestry between populations, recombination versus minor parent ancestry) at a series of genomic scales for each, with the smallest scale being the resolution of interpolation, and the largest scale corresponding to variation in signals occurring over roughly half of a chromosome. Wavelet correlations were averaged across chromosomes and error bars were obtained from a weighted jackknife procedure following [38]. To obtain the contribution of each scale to the overall correlation, we weighted correlations by the wavelet variances as described in [38]. We ran these analyses for interpolation distances of 1 kb and 32 kb.

Shared minor parent deserts and islands

We were interested in identifying regions that were likely under selection in both X. cortezi × X. birchmanni hybrid populations. Guided by the results of simulations, we used an ad hoc approach to identify regions with shared patterns of unusual ancestry across the 2 populations (see Text K in S1 File). We first identified ancestry informative sites where the minor or major parent ancestry fell in the lower 5% tail of genome-wide ancestry. We then selected the 0.05 cM window that overlapped this ancestry informative site and confirmed that the broader region fell within the lower 10% tail of genome-wide ancestry. We expanded out in the 5′ and 3′ directions in windows of 0.05 cM from this focal window until we reached a window on each edge that exceeded the 10% ancestry quantile. We treated this interval as an estimate of the boundary of the minor parent ancestry desert or island.

Because this approach may prematurely truncate ancestry deserts and islands (particularly in scenarios with error), in a separate analysis we merged any of deserts (or islands) that fell within 50 kb of each other. We filtered these merged regions to remove any regions with fewer than 10 ancestry informative sites, with fewer than 10 single nucleotide polymorphisms present in the recombination map, or that were less that 10 kb in length.

By defining deserts and islands in 0.05 cM windows, we could easily overlap these regions between different populations and determine how many are shared between sampling sites. This allowed us to define regions that have shared ancestry patterns between Chapulhuacanito and Santa Cruz despite their independent origin. To compare the observed number of shared ancestry deserts and islands to what we would expect by chance, given the overall patterns of ancestry variation along the genome in the 2 populations, we permuted the data in 0.05 cM windows and asked how frequently ancestry deserts and islands were identified as being shared in X. birchmanni × X. cortezi populations, as we had with the real data. We repeated this procedure 1,000 times. Based on these permutations, we found that few shared minor parent ancestry deserts or islands were expected by chance (Fig 4A).

Since ancestry in a given window is strongly correlated with ancestry in the neighboring windows, especially at smaller spatial scales, we also wanted to performed permutations that preserved this ancestry structure. Specifically, for Chapulhuacanito, we shifted the window labels of ancestry summarized in 0.05 cM windows by 12.5 cMs, and we asked whether any windows that were major (or minor) parent ancestry outliers in the shifted data overlapped with the ancestry deserts (or islands) identified in Santa Cruz (using the same criteria as in the real data). We repeated this procedure 132 times to fully tile the whole genome. Consistent with the naïve permutation approach described above, we found that few minor parent ancestry outliers in X. birchmanni × X. cortezi hybrid populations overlapped minor ancestry outliers identified in the other population by chance (S26 Fig).

We were interested in whether any of the shared deserts or islands between Chapulhuacanito and Santa Cruz were also ancestry outliers in X. birchmanni × X. malinche populations. Given the complexity of implementing simulations that would preserve LD structure across the 5 hybrid populations we wanted to evaluate, we simply performed the naïve simulations in 0.05 cM windows described above. Based on these permutations, we found that few minor parent ancestry outliers in X. birchmanni × X. malinche hybrid populations are expected to overlap minor parent deserts (or islands) identified in X. birchmanni × X. cortezi hybrid populations by chance.

Time series analysis

We were interested in understanding how ancestry at minor parent deserts and islands has changed over time. We focus this analysis on Chapulhuacanito due to insufficient sampling over time from the Río Santa Cruz (both in terms of numbers of hybrids sampled and number of sampling years available). The first samples we have access to from Chapulhuacanito are from 2003, approximately 40 generations ago. However, based on our demographic inference, this population likely underwent approximately 100 generations of evolution between initial hybridization and our first sampling year, meaning that even in our earliest samples we are evaluating ancestry in a late-stage hybrid population.

We focused our analysis on deserts and islands identified in 2021, but our results were qualitatively similar when ascertainment was performed in other years. Using the coordinates determined in 2021, we calculated average minor parent ancestry in the same region in each of the other years sampled (2002, 2006, and 2017). For minor parent islands, which showed greater levels of fluctuation in ancestry over time (see Results), we used a linear model implemented in R to test for a significant relationship between year and minor parent ancestry.

Ethics statement

Field sample collections were approved by Conapesca Permiso de Pesca de Fomento, permit No. PPF/DGOPA-002/19. Vertebrate animal sampling and experiments were approved by Stanford’s Administrative Panel on Laboratory Animal Care under protocol number 33071.