The role of gene flow in rapid and repeated evolution of cave-related traits in Mexican tetra, Astyanax mexicanus

2.1 Sampling, DNA extractions, and sequencing

We sampled five populations of Astyanax mexicanus cave and surface fish from the Sierra de Guatemala, Tamaulipas, and Sierra de El Abra region, San Luis Potosí, Mexico: Molino, Pachón, and Tinaja caves and the Rascón and Río Choy surface populations (Figure 1). Two main lineages of cave and surface populations are often referred to as “new” and “old” in reference to when the lineages presumably reached northern Mexico, each with independently evolved cave populations. Populations in Molino cave and Río Choy are considered “new” lineage fish, and the populations of Pachón cave and Tinaja cave are typically classified as “old” lineage populations (Bradic et al., 2012; Coghill et al., 2014; Dowling, Martasian, & Jeffery, 2002; Ornelas-García et al., 2008). These cave populations were the focus of our study because they are the most commonly used cave populations for laboratory studies.

Rascón was selected because cytochrome b sequences are most similar to old lineage cave populations (Ornelas-García et al., 2008). Rascón is part of the river system of Rio Gallinas, which ends at the 105 m vertical waterfall of Tamul into Rio Santa Maria/Tampaon. It is hypothesized that new lineage surface fish could not overcome the 105 m waterfall and could not colonize the Gallinas river (e.g., Rascón population). Thus, Rascón likely has remained cut-off from nearby surface streams inhabited by new lineage populations, and it, thus, was an important surface locality to include. Most other surface populations are thought to be of “new” origin, exhibit low divergence between each other, and be panmictic (Bradic et al., 2012); thus, sampling Río Choy is an initial proxy for sampling surface populations range-wide.

Fin clips were collected from fish in Spring 2013 from Pachón cave, Río Choy (surface), and a drainage ditch near the town of Rascón in San Luis Potosí, Mexico. Samples from Molino cave were collected by W. Jeffery, B. Hutchins and M. Porter in 2005, and Pecos drainage in Texas in 1994 by W. Jeffery and C. Hubbs. Samples from Tinaja cave were collected in 2002 and 2009 by R. Borowsky.

We complemented our sequencing of these populations with four additional individuals. First, we sequenced an A. mexicanus sample from a Texas surface population. While this population is not of direct interest to this study, it has been the focus of many Quantitative Trait Loci (QTL) studies (e.g., O'Quin et al., 2015), and therefore, it is evolutionary relationship to our populations is of interest. Second, we aimed to sequence an outgroup to polarize mutations in A. mexicanus. To do so, we first sequenced two samples of a close congener Astyanax aeneus (Mitchell et al., 1977; Ornelas-García & Pedraza-Lara, 2015) from Guerrero, Mexico. However, despite being separated from the sampled Astyanax mexicanus populations by the Trans-Mexican Volcanic Belt (Mitchell et al., 1977; Ornelas-García & Pedraza-Lara, 2015), we found a history of gene flow between Astyanax aeneus and Astyanax mexicanus. We therefore sequenced a more distant outgroup, the white skirt tetra (long-finned) Gymnocorymbus ternetzi. Gymnocorymbus ternetzi was used only for polarizing 2D site frequency spectra.

DNA was extracted using the Genomic-Tip Tissue Midi kits and DNEasy Blood and Tissue kit (Qiagen). We performed whole-genome resequencing with 100 bp paired-end reads on an Illumina HiSeq2000 at The University of Minnesota Genomics Center. Samples were prepared for Illumina sequencing by individually barcoding each sample and processing with the Illumina TruSeq Nano DNA Sample Prep Kit using v3 reagents. Five barcoded samples were pooled and sequenced in two lanes. In total, 45 samples were sequenced across 18 lanes. We resequenced nine Pachón cavefish, ten Tinaja cavefish, nine Molino cavefish, six Rascón surface fish, nine Río Choy surface fish and two Astyanax aeneus samples. We also made use of the Astyanax genome project to add another sample from the Pachón population, and obtain genotype information for the reference sequence. For this reference sample, we used the first read (R1) from all of the 100 bp paired-end reads that were sequenced using an Illumina HiSeq2000 and aligned to the reference genome (McGaugh et al., 2014). The Texas population and the white skirt tetra were barcoded, pooled and sequenced across two lanes of 125 bp paired-end reads from a HiSeq2500 in High-Output run mode using v3 reagents.

For all raw sequences, we trimmed and cleaned sequence data with trimmomatic v0.30 (Lohse et al., 2012) and cutadapt v1.2.1 (Martin, 2011) using the adapters specific to the barcoded individual, allowing a quality score of 30 across a 6 bp sliding window, and removing all reads <30 nucleotides in length after processing. When one read of a pair failed QC, its mate was retained as a single-end read for alignment. Post-processing coverage statistics were generated by the fastx toolkit v0.0.13 ( http://hannonlab.cshl.edu/fastx_toolkit/) and ranged from 6.8- to 11.9-fold coverage (mean = 8.87-fold coverage; median = 8.65-fold coverage Table 1, Supporting Information Tables S1 and S2), assuming a genome size estimate of 1.19 Gb (McGaugh et al., 2014) for the Illumina Hi-Seq 2000 samples. For the Pachón reference genome sequence, which was excluded from the measures above, the coverage was 16.68-fold. All new Illumina sequence reads were submitted to NCBI's Sequence Read Archive (Supporting Information Table S3). We expect some heterozygosity dropout with this relatively low level of coverage, but we sought to balance number of samples and depth of coverage in a cost-effective manner. Heterozygosity dropout would lower our diversity estimates, and if directional (e.g., due to reference sequence bias) could increase divergence between populations and lower estimates of gene flow. As detailed below, our data estimate more recent divergence times and higher estimates of gene flow than past studies.

2.2 Alignment to reference and variant calling

The Astyanax mexicanusgenome v1.0.2 (McGaugh et al., 2014) was downloaded through NCBI genomes FTP. Alignments of Illumina data to the reference genome were generated with the BWA-mem algorithm (Li, 2013) in bwa-0.7.1 (Li & Durbin, 2009, 2010). Both genome analysis toolkit v3.3.0 (GATK) and picard v1.83 ( http://broadinstitute.github.io/picard/) were used for downstream manipulation of alignments, according to GATK Best Practices and forum discussions (Auwera et al. 2013; DePristo et al., 2011; McKenna et al., 2010). Alignments of paired-end and orphaned reads for each individual were sorted and merged using Picard. Duplicate reads that may have arisen during PCR were marked using Picard's MarkDuplicates tool and filtered out of downstream analyses. GATK's IndelRealigner and RealignerTargetCreator tools were then used to realign reads that may have been errantly mapped around indels (insertions/deletions). Additional details are in Supporting Information Methods.

The haplotypecaller tool in gatk v3.3.0 was used to generate GVCFs of genotype likelihoods for each individual. The genotypegvcfs tool was used to generate a multi-sample variant call format (VCF) of raw variant calls for all samples. Hard filters were applied separately to SNPs and indels/mixed sites using the variantfiltration and selectvariants tools (Supporting Information Table S4). Filtering variants was performed to remove low confidence calls from the data set. The filters removed calls that did not pass thresholds for base quality, depth of coverage and other metrics of variant quality (Supporting Information Table S4).

Alleles with 0.5 frequency appeared to be overrepresented in the site frequency spectra, and the depth of coverage for alleles with 0.5 frequency was greater than the coverage at other frequencies. Upon examination, every individual was heterozygous and these alleles occurred in small tracts throughout the genome. We concluded these were likely paralogous regions (teleost fish have an ancient genome duplication, Hoegg, Brinkmann, Taylor, & Meyer, 2004; Meyer & Van de Peer, 2005). Molino was the most severely impacted population, likely due to Molino's low diversity allowing for collapsing of paralogs (see below). To be conservative, we identified sites where 100% of individuals in any population were heterozygotes and excluded these sites in all analyses. We still observed slight inflation at 0.5 frequency even after our heterozygosity filtering, and these are probably instances of collapsed paralogs where our criterion of heterozygosity in every genotyped individual is not met.

We generated a VCF file with variant and invariant sites, and subset it to include biallelic only SNPs where applicable. We excluded sites that were defined as repetitive regions from a Repeat Modeler analysis in McGaugh et al. (2014). Next, we removed sites that had less than six individuals genotyped in all populations. We also scanned for indels in the VCF file and removed the indel as well as 3 bp ± around the indel. In total, 171,841,976 bp were masked from downstream analyses. Lastly, all sites that were tri- or tetra-allelic were removed. Thus, sites were either invariant or biallelic. To reduce issues of non-independence induced by linkage disequilibrium, we analysed a thinned set of biallelic SNPs by retaining a single SNP in each non-overlapping window of 150 SNPs for the ADMIXTURE, TreeMix, and F₃ and F₄ analyses, resulting in approximately 119,000 SNPs for those analyses.

2.3 Population phylogeny and divergence

We generated a phylogenetic reconstruction to understand population relationships using 47 sampled individuals (excluding white skirt tetra, but including A. aeneus and Texas A. mexicanus). The population phylogeny estimation was implemented in raxml v8.2.8 with the two A. aeneus specified as the outgroups. We converted the VCF to fasta format alignments by passing through the mvf format using mvftools v3 (Pease & Rosenzweig, 2015). This was done to preserve ambiguity codes at sites polymorphic in individuals. Sites with greater than 40% missing data were removed using trimal v1.2 (Capella-Gutiérrez, Silla-Martínez, & Gabaldón, 2009). We performed 100 rapid bootstrapping replicates followed by an ML search from two separate parsimony trees. For computational efficiency, we implemented the GTRCAT model, though we note that inferences using a smaller subset of the genome and the GTRGAMMA model recover the same topology as presented in Figure 1. The ML tree with bootstrap support was drawn using the ape package v3.5 (Paradis, Claude, & Strimmer, 2004) in r v3.2.1 (R Core Team, 2018) using the A. aeneus samples as the outgroup; however, our results are consistent even if the outgroup is left unspecified and this decision is justified by the observation that A. aeneus is more distant from all A. mexicanus populations than any are to one another, as measured by the mean number of pairwise sequence differences.

2.4 Genomewide tests for gene flow

2.5 Demographic modelling

We modelled the demographic history of pairs of populations to further elucidate their relationships and migration rates. To conduct demographic modelling, we generated unfolded joint site frequency spectra (2D SFS) from the invariant sites VCF using a custom Python script. Detailed explanation of parameters and a user-friendly guide is available here: https://github.com/TomJKono/CaveFish_Demography/wiki.

White skirt tetra (Gymnocorymbus ternetzi) was used as the estimated ancestral state for all 10 pairwise combinations of Río Choy surface, Molino cave, Pachón cave, Rascón surface and Tinaja cave. Recently admixed individuals were excluded from the comparisons. To reduce the effect of paralogous alignment from the ancient teleost genome duplication, sites at 50% frequency in both populations were masked from the demographic modelling analysis. Sites were excluded if they were heterozygous in the representative outgroup sample or had any missing information in either of the populations being analysed. In addition, sites were excluded due to indels or repetitive regions as described above. Aside from these exclusions, all other sites, including invariant sites were included in all 2D SFS and the length of these sites make up locus length.

A total of seven demographic models were fit to each 2D SFS to infer the timing of population divergence, effective migration rates and effective population sizes using ∂a∂i 1.7.0 (Gutenkunst, Hernandez, Williamson, & Bustamante, 2009). The seven models were derived from previously published demographic modelling analysis (Tine et al., 2014). Briefly, the seven models are as follows: SI—strict isolation, SC—secondary contact, IM—isolation with migration, AM—ancestral migration, SC2M—secondary contact with two migration rates, IM2M—isolation with migration with two migration rates and AM2M—ancestral migration with two migration rates (see figure S9 of Tine et al., 2014). Two migration rates within the genome appeared to fit the data better in previous work (Tine et al., 2014), as this approach allows for heterogeneous genomic divergence (M which is most often the largest migration rate within the genome, Mi which is usually the lowest migration rate within the genome). Descriptions of the parameters and parameter starting values are given in Supporting Information Table S5 and Supporting Information Methods. For each replicate, Akaike information criterion (AIC) values for each model were converted into Akaike weights. The model with the highest mean Akaike weight across all 50 replicates was chosen as the best-fitting model (Rougeux, Bernatchez, & Gagnaire, 2017; Supporting Information Table S6; similar to Wagenmakers & Farrell, 2004).

The best-fitting demographic model for each 2D SFS comparison was used to generate estimates of the population parameters. Scaling from estimated parameters to real-value parameters was done assuming a mutation rate (μ) of 3.5 × 10⁻⁹ per bp per generation (Malinsky et al., 2017) and a locus length (L) of the number of sites used to generate the 2D SFS. For effective population size estimates, we estimated the reference population size, Nref, as the mean “theta” value from all 50 replicates of the best-fitting model multiplied by 1/(4 μL). The effective population sizes of the study populations were then estimated as a scaling of Nref. To estimate the per-generation proportion of migrants, we scaled the mean total migration rates from ∂a∂i by 1/(2Nref). Total times since divergence for pairs of populations were estimated as the sum of the split time (“Ts”) and time since secondary contact commenced (“Tsc”) estimates for each population pair.

2.6 Modelling of selection needed for cave alleles to reach high frequency

With demographic estimates provided by our whole-genome sequencing and ∂a∂i, we estimated the selection coefficients needed to bring alleles associated with the cave phenotype to high frequency. We implemented the 12-locus additive alleles model by Cartwright, Schwartz, Merry, and Howell (2017) with the parameters estimated for Molino cave, as this is the population where selection would need to be strongest to overcome the effects of drift. We estimated the one-way mutation rate of loci (μ) to be on the order of 1 × 10⁻⁶ (c.f. 3.5 × 10⁻⁹/bp/generation * 1170 bp, which is the median gene length across the genome). Other parameters were N = 7,335 (Molino Ne), h = 0.5 (additive alleles), k = 12 (number of loci) and Q = 0.1 (surface allele frequency of cave-favoured alleles). The simulation model was adapted from Cartwright et al. (2017) with the addition that the cave population was isolated for a period of time before becoming connected again with the surface. Consistent with our demographic models for Molino and Río Choy, migration rates spanned between 10⁻⁶ and 10⁻⁵ with 91,000 generations of isolation followed by 71,000 generations of connection. We also conducted simulations with the parameters from Tinaja cave as this is the cave population where selection would need to be weakest to overcome the effects of drift. All parameters were the same as above except N = 30,522 (Tinaja Ne), and 170,000 generations of isolation were followed by 20,000 generations of connection.

2.7 Candidate regions introgressed between caves

To identify regions of the genome that were likely transferred between cave populations or between cave and surface populations and were linked to cave-associated phenotypes, we implemented an outlier approach incorporating pairwise sequence differences and F_ST. To determine potential candidate regions, we utilized 5 kb d_XY windows using all biallelic sites, not just fourfold degenerate, since so few sites would be available. We first phased the biallelic variant sites for all samples using beagle (version 4.1; Browning & Browning, 2007). We then determined the number of invariant sites (from the full genome VCF) and variant sites in each 5 kb window to calculate average pairwise diversity. We then identified 5 kb regions in which d_XY was in the lower 5% tail of the genomic distribution of d_XY for all three pairwise combinations of cave populations, but substantial divergence with either surface population as measured by F_ST or d_XY. For F_ST outliers, we required that π per gene for both surface fish populations must be greater than the lowest 500 π values for genes across the genome. This requirement protects, in part, against including regions that are low diversity across all populations due to a feature of the genome.

To put the results in context of phenotypes mapped to the genome, we created a database of prior QTL from several key studies (Kowalko, Rohner, Linden et al., 2013; Kowalko, Rohner, Rompani et al., 2013; O'Quin, Yoshizawa, Doshi, & Jeffery, 2013; Protas, Conrad, Gross, Tabin, & Borowsky, 2007; Protas et al., 2008; Yoshizawa, Yamamoto, O'Quin, & Jeffery, 2012; Yoshizawa et al., 2015) and used overlapping markers between studies to position QTL relative to the linkage map in O'Quin et al. (2013). For markers in (Kowalko, Rohner, Rompani et al., 2013), we used blastn to place the marker on a genomic scaffold and placed the QTL from (Kowalko, Rohner, Rompani et al., 2013) in our database as locating to the entire scaffold. Our qtl database is given in Supporting Information Table S7.

3.1 Population phylogenetic reconstruction

We inferred the population phylogeny using half the nuclear genome (invariant and variant sites, but no mitochondrial sites) in the program raxml. Our phylogenetic tree clearly demarcates two lineages and indicates that the Rascón surface population and the Tinaja and Pachón cave populations form a monophyletic clade (Figure 1; often referred to as “old lineage”). Similarly, the Río Choy surface and Molino cave populations form a monophyletic clade (referred to as “new lineage”). Thus, this phylogeny indicates that many cavefish traits may be polyphyletic (i.e., evolved through repeated evolution).

In agreement with a previous study showing that Molino cave population has the lowest diversity of cave populations sampled, the Molino cave population exhibits shorter branches than other populations tested (Bradic et al., 2013), while the surface populations exhibit longer branch lengths. While all populations and the two lineages have high bootstrap support (≥99), bootstrap support with genomic-scale data provides little information about the evolutionary processes or the distribution of alternative topologies across the genome (Yang & Rannala, 2012); thus, we use a series of tests below to explore this further.

3.2 Diversity and divergence

Patterns of diversity within populations, π, and interpopulation divergence, d_XY, provide a broad summary of the coalescent history within and between populations. We present pairwise sequence differences at fourfold degenerate sites between all pairs of individuals (Figure 2). The “striped” individuals in Figure 2b correspond to recent hybrids as inferred by ADMIXTURE (Figure 2a). We removed these putative recent hybrids from our summary of diversity within and divergence between populations presented in Table 2.

Genomewide average diversity within cave populations (π_{4fold degen}: Molino = 0.00074, Tinaja = 0.00129, Pachón = 0.00100; Table 2) is substantially lower than diversity within surface populations (π_{4fold degen}: Río Choy = 0.00300, Rascón = 0.00207), reflecting a decrease in effective population size in caves (sensu Avise & Selander, 1972). Molino cave is the most homogenous. These results are consistent with the short branches in cave populations observed in the phylogenetic tree produced by RAxML (Figure 1). In contrast, the surface population Río Choy is the most diverse in our sample—so much so that two Río Choy fish may be more divergent than fish compared between any of the old lineage populations (e.g., Tinaja cave and Rascón sruface).

Genomewide average divergence between the new and old lineages (d_{XY 4fold degen} = 0.00340 − 0.00374, Table 2) exceeds divergence between cave and surface populations within lineages (d_{XY 4fold degen} Molino-Río Choy = 0.00332; Tinaja-Rascón = 0.00285; Pachón-Rascón = 0.00298; Table 2). Thus, old and new lineages diverged prior to (or have experienced less genetic exchange than) any of the cave–surface population pairs within lineages. The two old lineage caves are the least diverged populations in our sample (d_{XY 4fold degen}: Pachón-Tinaja = 0.00225). These results are consistent with both the observation of monophyly of old and new lineages and the observation that old lineage cave populations are sister taxa in the raxml tree (Figure 1).

Divergence between populations suggests that a simple bifurcating tree does not fully capture the history of these populations. For example, divergence between the new lineage cave population (Molino) and the geographically closer old lineage cave population (Pachón) is less than divergence between Molino and the geographically further old lineage populations (Tinaja and Rascón; Table 2). Likewise, all old lineage populations are closer to Molino cave than they are to Río Choy. These results suggest gene flow between Molino cave and the old lineage populations, which is supported by additional analyses.

3.3 Genomewide tests suggest substantial historical and contemporary gene flow

3.3.2 F₃ and F₄ statistics

To test for historic genetic exchange, we calculated F₃ and F₄ “tree imbalance tests,” after removing recent hybrids identified by ADMIXTURE. Throughout the section below, gene flow is supported by F₄ statistics between the two underlined taxa and/or the two non-underlined taxa in the four-taxa tree. Further, we interpret the F₄ statistics in relation to the ranking of the F₄ scores. Extreme F₄ scores may be the result of both pairs (e.g., A–C and B–D) exchanging genes, amplifying the F₄ score beyond what the score would be if gene flow occurred only between a single pair within the quartet.

F₃ and F₄ tests convincingly show that Rascón, the old lineage surface population, has experienced gene flow with a lineage more closely related to A. aeneus than to the other individuals in our sample (Table 3). This claim is supported by numerous observations including the observations that, while all F₃ tests including Rascón are significantly negative (i.e., supporting admixture), the most extreme F₃ statistic is from a test with Rascón as the target population and A. aeneus and Tinaja as admixture sources, reflecting that both A. aeneus and Tinaja have likely hybridized extensively with Rascón. Additionally, all F₄ tests of the form (A. aeneus, new lineage; old lineage cave, Rascón) are significantly negative, meaning that Rascón is genetically closer to A. aeneus than expected given a simple tree (Supporting Information Figure S3). Therefore, A. aeneus is an imperfect outgroup to A. mexicanus. Our interpretation is that A. aeneus and Rascón hybridized too deep in the past to be detected by ADMIXTURE analysis, but sufficiently recently to be detected by F₃ and F₄ tests. This is further supported by an analysis included in the supplemental materials that did not find long blocks of sequence similarity to A. aeneus in the Rascón genome.

Table 3. F₃ statistics for significant configurations out of all possible three population configurations (X; A, B) where X is the population tested for admixture. A significantly negative F₃ statistic indicates admixture from populations related to A and B. Notably, we conducted the F₃ statistics without individuals showing recent evidence of admixture. Sample sizes were as follows: A. aeneus: N = 2; Río Choy: N = 8; Molino N = 9; Rascón N = 5; Pachón N = 9; Tinaja N = 8. Any z-score below −1.645 passes the critical value for significance at α = 0.05. Tests are ordered most-least extreme z-scores. All other confirmations were not significant for introgression

(X; A, B)	F3-statistic	SE	z-score
Rascón; A. aeneus, Tinaja	−9.25e-04	5.63e-05	−16.42
Rascón; Río Choy, Tinaja	−7.34e-04	4.97e-05	−14.78
Rascón; Molino, Tinaja	−7.21e-04	5.17e-05	−13.94
Rascón; A. aeneus, Pachón	−5.52e-04	6.05e-05	−9.13
Rascón; A. aeneus, Molino	−3.73e-04	6.53e-05	−5.71
Rascón; Río Choy, Pachón	−2.04e-04	6.03e-05	−3.38
Rascón; Molino, Pachón	−1.46e-04	6.18e-05	−2.36
Río Choy; A. aeneus, Molino	−1.02e-04	5.51e-05	−1.84

Our results also show that the new lineage surface population, Río Choy, experienced gene flow with an unsampled outgroup. A significantly negative F₃ value demonstrates Río Choy (the new lineage surface population) is an admixture target with Molino (new lineage cave) and A. aeneus as admixture sources (Table 3). This claim is bolstered by significantly positive F₄ values in all comparisons of the form (A. aeneus, old lineage; Río Choy, Molino) in Supporting Information Figure S3. However, because d_XY between A. aeneus and both new lineage populations are nearly equivalent, we suggest this is an example of “the outgroup case” for F₃ (Patterson et al., 2012) in which admixture is attributed to an uninvolved outgroup (in this case, A. aeneus) rather than the unsampled admixture source.

We uncover evidence for admixture between the old lineage cave population, Pachón, and both new lineage populations (Molino cave and Río Choy surface), as observed in recent hybrids in ADMIXTURE. Specifically, the significantly negative F₄ values for (A. aeneus, new lineage; Pachón, Tinaja) (Supporting Information Figure S3) are consistent with gene flow between both new lineage populations and the old lineage Pachón cave, and/or gene flow between Tinaja and A. aeneus. Because no other evidence suggests gene flow between Tinaja and A. aeneus, we argue that this result reflects gene flow between Pachón and the new lineage populations. This claim is further supported by the observation that d_XY between new lineage populations and Pachón is consistently lower than d_XY between new lineage populations and the other old lineage populations (Table 2).

Additionally, our analyses are consistent with gene flow between the old lineage Tinaja cave and the old lineage surface population, Rascón. Again, this result complements the discovery of two recent hybrids between these populations in our ADMIXTURE analysis. This claim is supported by the significantly positive F₄ value for (A. aeneus, Rascón; Pachón, Tinaja). Since we lack evidence of Pachón – A. aeneus admixture, this significant F₄ statistic is likely driven by Rascón – Tinaja admixture.

Despite gene flow among most population pairs, there is one case in which tree imbalance tests failed to reject the null hypothesis of a bifurcating tree: F₄ tests of the form (new lineage, new lineage; old lineage, old lineage) (Supporting Information Figure S3). This result is counter to both our observation of recent hybrids (between the new and old lineages observed in ADMIXTURE), and the gene flow observed in other F₄ – statistics, as well as treemix (Figures 3, Supporting Information Figure S4), d_XY nearest neighbour proportions (Supporting Information Figures S5–S7) and phylonet (Supporting Information Figure S8). In these cases, it is also plausible that F₄ scores that do not differ from zero may reflect opposing admixture events that cancel out a genomewide signal, rather than an absence of introgression (see Reich et al., 2009).

3.4 Divergence times are younger than expected without accounting for migration

The levels of diversity and divergence allow us to calculate a simple estimate of population split times. This approach results in remarkably recent divergence time estimates. For example, we estimate the oldest split between old and new lineages of approximately 156,597 generations (Table 2; using π_A. aeneus as a proxy for ancestral diversity, and representing the old-new lineage split by d_{XY Choy,Rascón}, ((4.68 − 3.58) × 10⁻³)/ (7 × 10⁻⁹). Assuming the generation interval of 1 year, the oldest estimate for the old and new split is more than an order of magnitude less than that estimated from the cytochrome b which places the divergence time between lineages between 5.7 and 7.5 mya (Ornelas-García et al., 2008). Similarly, we estimate that the split between new lineage cave (Molino) and new lineage surface (Río Choy) populations was approximately 113,913 generations ago ((4.38 − 3.58) × 10⁻³)/(7 × 10⁻⁹) similar to a recent analyses (Fumey et al., 2018). However, other estimates using this method were unstable, suggesting that A. aeneus is not a good proxy for ancestral diversity in the old lineage and/or introgression obscures any realistic estimate of divergence time. Our results that suggest gene flow (even with the exclusion of recently admixed individuals) indicate that true divergence times exceed those estimated from comparisons of d_XY and π. We, therefore, pursue a model-based estimate of population divergence time while accounting for introgression.

3.5 Demographic modelling indicates cave populations are younger than expected when accounting for migration

Demographic modelling using ∂a∂i revealed extensive interdependence and contact for all populations studied. In all cases, models with multiple migration rates fit the population comparisons best, suggesting that some genomic regions are more recalcitrant to gene flow than others. For most pairwise population comparisons, the best-fitting models supported a period of isolation followed by a period of secondary contact (SC2M; Supporting Information Table S6). The only exceptions were the comparisons between Molino-Rascón and Pachón-Rascón that supported divergence with isolation (IM2M) slightly more than SC2M (Supporting Information Table S6). The distributions of divergence times for IM2M were multimodal (Molino-Rascón) or nearly flat (Pachón-Rascón) (Supporting Information Figure S9); thus, we present results from SC2M models that exhibit much tighter distributions for divergence estimates (Table 4; Supporting Information Figure S9).

Effective population sizes for the surface populations were an order of magnitude higher than for the cave populations, and Molino exhibited the lowest effective population size of all five populations (Supporting Information Table S8). Notably, even for cave populations, estimates of effective population size were an order of magnitude larger than previous estimates (Avise & Selander, 1972; Bradic et al., 2012) and on par with mark-recapture census estimates (~8,500 fish in Pachon cave with wide 95% confidence interval of 1,279–18,283; Mitchell et al., 1977).

∂a∂i demographic modelling estimated deeper split times for populations than the divergence estimates calculated as T = (d_XY − π_ancestral)/2μ (above) because they account for introgression that can artificially deflate these divergence estimates. The oldest divergence time between old and new lineages was estimated at 256,675 generations before present (Río Choy-Rascón), (Table 4, Supporting Information Tables S9 and S10; Figure S9; using SC2M models for all comparisons). Importantly, both the model-independent method and the demographic estimates presented here reveal relatively similar divergence times between the old and new lineages (~157 vs. ~257 k generations ago), and these estimates differ substantially from the several million year divergence time obtained through mtDNA (Ornelas-García et al., 2008), unless we have dramatically underestimated generation interval in the wild.

Cave–surface splits for both the new and old lineage comparisons were remarkably similar (Pachón – Rascón: 161,262 generations; Tinaja – Rascón: 190,650 generations; Molino – Río Choy: 162,537 generations). While the two old lineage cave populations (Pachón – Tinaja) are estimated to have split slightly more recently (115,522 generations before present), suggesting that colonization of one of the two old lineage caves was potentially from subterranean gene flow (as suggested by Espinasa & Espinasa, 2015). Distributions and estimates across 50 replicates are given in Supporting Information Figure S9, Tables S9 and S10.

Subterranean gene flow may be higher than surface to cave or between-lineage surface gene flow. The highest rates of gene flow are estimated between caves, and surprisingly, Molino cavefish exhibited a migration rate into Tinaja cave that is higher than Tinaja gene flow into Pachón cave. Surface fish gene flow rates into cave populations are also among the highest rates (especially Río Choy into all sampled caves). Notably, several caves also have relatively high rates of gene flow into surface populations (Molino into Río Choy and Tinaja into Rascón).

Heterogeneity in gene flow across the genome is also different across population pairs. Generally, more of the genome of cave–surface pairs (48%) seems to follow the lower migration rate than in cave–cave pairs (40%) and surface–surface pair (18%), suggesting the possibility that strong selection for habitat-specific phenotypes slows gene flow across the genome.

3.6 Modelling of selection needed for cave alleles to reach high frequency

Similarly to that estimated by (Cartwright et al. 2017), selection coefficients of ~0.01 are needed for cave phenotype alleles to reach high frequencies in either Molino or Tinaja cave populations with a 12-locus additive model (Figure 4). Animations of allele frequencies at different levels of selection across the historical demography are provided in Supplementary materials. There is an increase in noise at the secondary contact period because the influx of new alleles. This has little effect overall because the migration rate is of the same magnitude as the mutation rate and may be because prior to secondary contact, the loci are typically fixed for either the cave or surface allele and after contact, the loci may become more polymorphic. The average selection coefficients appear to approximately match the pre-secondary contact average.

In sum, our data support that cavefish population sizes are sufficiently large for selection to play a role shaping traits, gene flow is sufficiently common to impact repeated evolution, and cave–surface divergence is more recent than expected from mitochondrial gene trees.

3.7 Candidate gene regions introgressed between caves

In light of the suspected gene flow among populations, we identified genomic regions that showed molecular evolution signatures consistent with gene flow between caves and genomic regions that are likely affected by cave–surface gene flow. There were 997 5 kb regions (out of 208,354 total) that exhibited the lowest 5% of d_XY windows of the genome for each of the cave–cave comparisons (Pachón-Molino, Pachón-Tinaja, Molino-Tinaja). These windows were spread across 471 scaffolds and overlapped 500 genes. We highlight some of these 500 genes in the context of Ensembl phenotype data in Supplementary text (see also Supporting Information Table S11).

Several of these regions cluster with known QTLs (O'Quin & McGaugh, 2015). We examined co-occurring QTLs for many traits on Linkage Group 2 and Linkage Group 17 (LG2, LG17, Supporting Information Table S12), which were two main regions highlighted in recent work on cave phenotypes (Yoshizawa, Yamamoto et al., 2012). We found that the co-occurring QTL on Linkage Group 2 harboured about 1.5-fold (odds ratio = 1.54, 95% CI = 1.139–2.089) the number of regions with low divergence among all three cave populations (44 windows with all three caves in lowest 5% of d_XY/6070 total windows that had data across these same scaffolds) relative to the total across the entire genome (997/208354). This suggests that this region with many co-occurring QTL was potentially transferred among caves. Notably, we did not find a similar pattern for linkage group 17 (proportion of windows across all three caves in 5% lowest d_XY = 0.00482 for LG17 and 0.00479 across the entire genome). The region on Linkage Group 2 is not simply an area of high sequence conservation (which could account for low divergence among all three cave populations) because the mean divergence between surface-surface comparisons for this region is very similar to the genomewide average (genomewide d_XY = 0.00481, LG2 QTL region d_XY = 0.00482).

We find genes that look to be affected by gene flow between caves. We found 1004 genes that exhibited d_XY = 0 and another 370 genes with d_XY in the lowest 5% across the genome for the three pairwise comparisons of the caves, many with substantial divergence to at least one surface population (Supporting Information Table S13). We associated these genes with previously known QTL with linkage groups that follow (O'Quin et al., 2013) and a qtl database is provided (Supporting Information Table S7). One example is fam136a (family with sequence similarity 136, member A), which is under a QTL for body condition on Linkage Group 10. This gene is expressed in the hair cells of the crista ampullaris, an organ important for detecting rotation and acceleration, in the semicircular canals of the inner ear in the rat (Requena et al., 2014). All caves are identical except for a SNP that may have come from the Rascon population in one of the admixed Tinaja individuals. fam136a is in the top 10% most divergent genes by d_XY for Rascón surface—Pachón cave and Rascón surface—Tinaja cave populations (Molino-Rascón surface comparisons fall in the top 13% most divergent genes via d_XY). Notably, Rascon and A. aeneus exhibit an E (glutamic acid) at 15aa (suggesting this is the ancestral state), whereas Río Choy and all caves exhibit a G (glycine). This amino acid switch is between a polar and a hydrophobic, thus not very functionally conserved. Such pattern could be produced by gene flow among caves or by gene flow of Río Choy with the cave populations.

4.1 Cave populations are younger than previously estimated

Our estimates fit in well with previous suggestions that the caves were colonized in the late Pleistocene (Avise & Selander, 1972; Fumey et al., 2018; Porter et al., 2007; Strecker et al., 2004) with both our d_XY-based divergence estimation and demographic modelling. As our work documents extensive gene flow, we favour the divergence time estimates provided by demographic models that incorporated estimates of migration. Demographic models estimate split time between the two lineages at approximately 257 k generations ago, and estimates of cave–surface population splits are ~161–191 k generations ago. Notably, Astyanax is a species of fish with a limited distribution in northern latitudes, with its current most northern locality in the Edwards Plateau in Texas (Page & Burr, 2011). These split times are consistent with cooler temperatures in Northern Mexico associated with glaciation playing some role in the colonization of the caves by Astyanax mexicanus, potentially as thermally stable refugia (Cussac, Fernández, Gómez, & López, 2009). Our nuclear genomic data suggest that the “old” and “new” lineage split is more recent than the main volcanic activity of the Trans-Mexican Volcanic Belt (3–12 Mya) which was thought to have separated these two lineages, as well as other lineages in other vertebrates (Ornelas-García et al., 2008). This geographic barrier likely has been breached by Astyanax multiple times and likely led to cave invasions with each migration (Gross, 2012; Hausdorf et al., 2011; Strecker et al., 2012).

4.2 Historic and contemporary gene flow between most populations

Several cave populations exhibit intermediate phenotypes and experience flooding during the rainy season (Espinasa, unpublished, Strecker et al., 2012), suggesting that intermediate phenotypes are the result of admixture between cavefish and surface fish swept into caves during flooding (Avise & Selander, 1972; Bradic et al., 2012). However, it has been suggested that surface fish and hybrids are too maladapted to survive and spawn in caves (Coghill et al., 2014; Hausdorf et al., 2011; Strecker et al., 2012), and that cavefish populations with intermediate troglomorphic phenotypes represent more recent cave invasions, rather than hybrids (Hausdorf et al., 2011; Strecker et al., 2012). Despite evidence from microsatellites (Bradic et al., 2012; Panaram & Borowsky, 2005), mitochondrial capture (Dowling et al., 2002; Ornelas-García & Pedraza-Lara, 2015; Yoshizawa, Ashida, & Jeffery, 2012), and haplotype sharing of candidate loci (but see Espinasa, Centone, & Gross, 2014; Gross & Wilkens, 2013), there was still uncertainty in the literature regarding the frequency of gene flow between Astyanax cavefish and surface fish and the role gene flow may play in adaptation to the cave environment (Coghill et al., 2014; Espinasa & Borowsky, 2001; Hausdorf et al., 2011; Strecker et al., 2012).

Our work demonstrates recent and historical gene flow between cave and surface populations both within and between lineages (Figures 2 and 3). This result is also suggested by past work which showed most genetic variance is within individuals (Table 2 in (Bradic et al., 2012)) and very few private alleles (i.e., alleles specific to a population) were present among cave populations (Figure 3b in (Bradic et al., 2012)). Notably, many of the methods we employed take into account incomplete lineage sorting, are robust to nonequilibrium demographic scenarios, and detect recent admixture (<500 generations ago) (Patterson et al., 2012). In all our analyses the hybridization detected may not be between these sampled populations directly, but unsampled populations/lineages related to them (Patterson et al., 2012).

One of the most intensely studied cave populations is Pachón. Our data of Pachón cavefish hybridization with the new lineage are expected from past field observations and molecular data. Only extremely troglomorphic fish were found early (1940s–1970s; Avise & Selander, 1972; Mitchell et al., 1977) and late (1996–2000, (Dowling et al., 2002)) surveys, though, phenotypically intermediate fish were observed in 1986–1988, as well as in 2008 (Borowsky, unpublished). Thus, subterranean introgression from a nearby cave population may cause transient complementation of phenotypes or hybridization with surface fish (from flooding or human-introduction) may contribute to the presence of intermediate-phenotype fish in the cave (Langecker, Wilkens, & Junge, 1991; Wilkens & Strecker, 2017). Indeed, past studies suggested gene flow between Pachón and the new lineage populations, as Pachón mtDNA clusters with new lineage populations (Dowling et al., 2002; Ornelas-García & Pedraza-Lara, 2015; Ornelas-García et al., 2008; Strecker et al., 2003, 2012). Interestingly and divergent from past studies, Pachón mitochondria group with old lineage populations in the most recent analysis (Coghill et al., 2014). Continued intense scrutiny of other caves may reveal similar fluctuations in phenotypes and genotypes.

The most surprising signal of gene flow in our data is seen in the ∂a∂i analyses, in which the signal of Molino-Tinaja exchange is similar to that from Pachón-Tinaja, suggesting subterranean gene flow between all three caves, despite substantial geographic distances (the entrances are separated by >100 km; Mitchell et al., 1977). While our genomewide ancestry proportion data suggest some exchange of Molino with Tinaja (Supporting Information Figure S5–S7), we have no other evidence to corroborate this signal. Thus, the results from population demography could be picking up a signal of Molino with Tinaja that is actually driven by Pachón—new lineage hybridization and subsequent Tinaja—Pachón hybridization or Tinaja—Río Choy hybridization. Also, surprisingly, ∂a∂i analyses suggest that some of the highest migration occurs from Tinaja cave into Rascón surface and Molino cave into Río Choy surface.

Notably, many species of troglobites other than Astyanax inhabit caves spanning from southern El Abra to Sierra de Guatemala, and some were able to migrate among these areas within the last 12,000 years (Espinasa, Bartolo, & Newkirk, 2014). Due to the extensive connectivity and the span of geological changes throughout the hydrogeological history, it is likely that there have been ample opportunities for troglobites, including Astyanax, to migrate across much of the El Abra region (Espinasa & Espinasa, 2015).

Across many systems genomic data has revealed reticulate evolution is much more common than previously thought (Abbott, Barton, & Good, 2016; Arnold & Kunte, 2017; Brandvain et al., 2014; Dasmahapatra et al. 2012; Geneva, Muirhead, Kingan, & Garrigan, 2015; Malinsky et al., 2017; McGaugh & Noor, 2012; Rougemont et al., 2017), and the ability of populations to maintain phenotypic differences despite secondary contact and hybridization is increasingly appreciated (Arnold & Kunte, 2017; Fitzpatrick, Gerberich, Kronenberger, Angeloni, & Funk, 2015; Malinsky et al., 2017; Payseur & Rieseberg, 2016). Thus, despite gene flow from the surface into caves, it is entirely possible that cave-like phenotypes can be maintained. Indeed, gene flow may help to sustain cave populations from the effects of inbreeding (Åkesson et al., 2016; Ellstrand & Rieseberg, 2016; Fitzpatrick et al., 2016; Frankham, 2015; Kronenberger et al., 2017; Whiteley, Fitzpatrick, Funk, & Tallmon, 2015) or catalyze adaptation to the cave environment (sensu Clarkson et al., 2014; Meier et al., 2017; Richards & Martin, 2017). Though the rates of gene flow between Astyanax populations is considerably less than that documented in other contemporary diverging fish lineages (coastal and marine anchovies, Le Moan, Gagnaire, & Bonhomme, 2016; parasitic river and nonparasitic brook lampreys, Rougemont et al., 2017; benthic and dwarf limnetic lake whitefish, Rougeux et al., 2017; Atlantic and Mediterranean sea bass, Tine et al., 2014), they are similar to the historic gene flow between of polar bears into brown bears (Liu et al., 2014). Our data indicate hybridization between sampled lineages and suggest that cavefish are poised to be a strong contributor to understanding the role gene flow may play in repeated evolutionary adaptation.

4.3 Candidate genes shaped by gene flow

In some cases, we have evidence that particular alleles were transferred between caves, but are highly diverged from surface populations (Supporting Information Table S11 and S13). This suggests that gene flow between caves may hasten adaptation to the cave environment and suggests that repeated evolution in this system may, in part, rely on standing genetic variation. Notably, scaffolds under the co-localizing QTL on Linkage Group 2 for many traits exhibit 1.5-fold enrichment for 5 kb regions with lowest divergence across caves relative to genomewide, suggesting parts of this QTL have spread throughout caves.

For genes with high similarity among caves and substantial divergence between cave and surface populations, many appeared to be involved in classic cavefish traits such as pigmentation and eye development/morphology (mdkb, atp6v0ca, ube2d2 l, usp3 and cln8) and circadian functioning (usp2a) (Supporting Information Table S13). Additional common annotations suggest future areas of trait investigation, namely, cardiac-related phenotypes (mmd2, cyp26a1, tbx3a, tnfsf10, alx1 and ptgr1) as well as inner ear phenotypes (ncs1a, dlx6a, fam136a) (Supporting Information Table S13). Importantly, the sensory hair cells of the inner ear are homologous with sensory hair cells of the neuromasts of the lateral line (Fay & Popper, 2000); therefore, these genes may be impacting mechanoreception or sound reception.

4.4 Effective population sizes are much larger than expected and weak selection can drive cave phenotypes

Previous estimates of very small effective population sizes in cave populations of A. mexicanus suggested drift and relaxed selection shaped cave-derived phenotypes (e.g., Lahti et al. 2009). The estimates of nucleotide diversity and N_e provided here (Tables 1 and 2 and Supporting Information Table S8) indicated that positive selection coefficients need not be extreme to drive cave-derived traits (Akashi, Osada, & Ohta, 2012; Charlesworth 2009). To put these values in perspective, the cave populations have an average genetic diversity (which is proportional to N_e) somewhat similar to humans and the surface populations exhibit a similar genetic diversity to zebrafish (Leffler et al. 2012).

Our observations support theoretical and empirical results that selection likely shaped cavefish phenotypes (Borowsky, 2015; Cartwright et al. 2017; Moran, Softley, & Warrant, 2014). Simulations using demographic parameters from Molino and Tinaja cave populations suggest that selection coefficients across 12 additive loci (patterned after the number of eye-related QTL (O'Quin & McGaugh, 2015)) need to be above 0.01 to bring cave alleles to very high frequencies (Figure 4). Such selection coefficients are often found driving selective sweeps in natural systems (Nair et al., 2003; Rieseberg & Burke, 2001; Schlenke & Begun, 2004; Wootton et al., 2002).

4.5 Diverging lineages should be examined with reticulation methods and absolute divergence metrics

Our genomewide data allowed a unique perspective that was not available in past studies. First, our data represent an empirical demonstration of how well-supported phylogenetic reconstructions can be misleading. When using genome-scale data with maximum likelihood, bootstrap values are not a measure of the number of sites that support a particular phylogeny, though, this is often how they are interpreted (Yang & Rannala, 2012). Rather, if a genome-scale data set is slightly more supportive of a particular topology, maximum likelihood will find that topology consistently and exhibit high levels of bootstrap support (Yang & Rannala, 2012). With the cavefish, a previous RADseq study suggested relatively strongly supported branches (Coghill et al., 2014); however, a much more complex evolutionary history with reticulation among lineages was revealed here. Examination of recently diverged taxa with reticulation methods (e.g., Phylonet, F₄, F₃, TreeMix) may ensure a more comprehensive view of their evolutionary history than the successive bifurcations represented in a tree.

Second, our data set is yet another empirical demonstration of poor performance of pairwise F_ST in estimating population relationships when diversity is highly heterogeneous among populations (Figure 5, Table 2, Supporting Information Table S14; Fumey et al., 2018; Jakobsson, Edge, & Rosenberg, 2013). When diversity is highly heterogeneous, this can give the false impression that low-diversity populations are highly divergent, when in reality, the lower diversity drives greater pairwise F_ST, not higher absolute divergence (Charlesworth, 1998). These limitations of F_ST are especially important to appreciate in systems like the cavefish where diversity is highly heterogeneous across populations. Further, it is often suggested that high F_ST translates to low gene flow, but violations of the assumptions are common (Cruickshank & Hahn, 2014). In the case of Molino (a very low diversity population), high F_ST values were taken to indicate relative isolation from other populations with little gene flow (Bradic et al., 2012), and we show here that is not the case. Indeed, pairwise F_ST among surface fish populations is the lowest and pairwise F_ST among caves is the highest, yet absolute divergence between surface populations is slightly higher than absolute divergence among caves (Figure 5, Table 2 and Supporting Information Table S14). Future molecular ecology work should assay diversity and interpret pairwise F_ST accordingly or use absolute measures of divergence in conjunction with pairwise F_ST (Charlesworth, 1998; Cruickshank & Hahn, 2014; Noor & Bennett, 2009; Ritz & Noor, 2016).