Epas1 genotype varies with altitude in deer mice

In order to examine altitudinal patterns of allele frequency variation of the genes encoding HIFs, and to put these patterns into a broader genomic context, we sequenced the exomes of 37 lowland mice from Lincoln, NE (430 m a.s.l.), and 48 highland mice from Mt. Evans, CO (4350 m a.s.l). Fifteen mice from a lowland population in Merced County, CA (~320 m a.s.l.), were included to infer polarity of DNA changes in highland mice. All exomes were sequenced using a custom Nimblegen probe set targeting exons from 25,246 nuclear genes (see Materials and Methods). Captured exomes were paired-end sequenced on an Illumina HiSeq 4000 and mapped to a reference genome (NCBI GCA_000500345.1 Pman_1.0). The final set of quality-filtered sites consisted of 5,182,530 high-quality bi-allelic variants sequenced at approximately 18X coverage (S1 Fig). Analyses of population genetic structure (using PCA [40] and Admixture [41]), revealed that all three populations were genetically distinguishable (S2 Fig and S3 Fig). Pairwise F_ST values (estimated with Weir’s Theta [42]) between Mt. Evans and Lincoln were 0.025± 3.16e-5 (mean±SEM), between Mt. Evans and Merced were 0.025±6.54e-5, and between Lincoln and Merced were 0.044±8.00e-5.

Based on these results, we calculated the population branch statistic (PBS [13]) for each single nucleotide polymorphism (SNP) to identify variants that exhibit extreme allele frequency changes in the highland population (Mt. Evans) relative to both lowland populations (Lincoln and Merced) (S4 Fig). Among the upper 0.1% of the PBS distribution, the only SNPs located in HIF genes were three SNPs located in the HIF-2α gene Epas1 (Fig 1A, S5 Fig); one of these SNPs was located in the 3’UTR, one was a non-synonymous polymorphism located at site 755 in the 14^th exon that changed threonine to methionine (^Thr755^Met), and one was a synonymous polymorphism also located in the 14^th exon. The highest-ranking SNP of these three was the non-synonymous, polarity-altering ^Thr755^Met polymorphism (PBS upper 0.1%; Fig 1A). Due to significant linkage disequilibrium between alleles at the three closely linked Epas1 SNPs (Fig 1A), and because there were no SNPs in any of the genes that encode HIF-1α or HIF-3α in the upper 0.1% of the empirical PBS distribution, we focused our subsequent analyses on the ^Thr755^Met mutation in Epas1.

Download:

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

Fig 1. Epas1 is an exome-wide outlier under spatially varying selection in P. maniculatus along altitudinal gradients.

A) Manhattan plot of PBS values for all SNPs (black dots) located within the last three exons of Epas1 (exon numbers provided above schematic). Exome-wide values for mean, top 1%, and top 0.1% percentile PBS values are shown, and three outlier SNPs in Epas1 are highlighted in orange (see key). Pairwise linkage disequilibrium estimates (measured with the squared correlation coefficient, r²) for each SNP pair are provided. B) Geographic variation in ^Thr755^Met Epas1 allele frequency for 23 populations in the Rocky Mountains and Great Plains, USA. Pie charts are shaded according to frequency of high-altitude or low-altitude allele, with size indicating number of mice sampled (see key). C) Clinal variation in ^Thr755^Met Epas1 allele frequency for 10 P. maniculatus populations sampled along a 4500 m altitudinal cline from the Great Plains of Nebraska to the Rocky Mountains in Colorado. In (B) and (C), Mt. Evans (ME) and Lincoln (LN) populations are labeled. Dashed box in (B) shows populations chosen for assessing clinal variation in (C). See S1 Table for details on sampling location and Epas1 allele frequencies.

https://doi.org/10.1371/journal.pgen.1008420.g001

To more broadly assess the relationship between Epas1 ^Thr755^Met and elevation, we genotyped an additional 266 deer mice collected from 23 sites across the western U.S. (S1 Table). We found that the Met-755 allele (henceforth called the Epas1^H allele) is significantly and positively correlated with altitude (r² = 0.589, p<0.001; Fig 1B; S6 Fig). For a single altitudinal transect connecting Lincoln to Mt. Evans, variation in Epas1 allele frequency is best explained as a sigmoidal cline centered at 1399.5 m a.s.l. (95% CI 1192.99–1493.01 m a.s.l.) (Fig 1C). Notably, the Epas1 cline is similar in shape, width, and center to that of ß-globin [43] (S7 Fig), a locus known to be under selection in high-altitude deer mice [28,43,44]. To infer character polarity of the amino acid change, we genotyped mice from nine additional Peromyscus species, including P. keeni, P. melanotis, P. hylocytes, P. attwateri, P. melanophrys, P. eremicus, P. polionotus, P. leucopus, as well as an outgroup rodent species, Reithrodontomys montanus. This broader phylogenetic sampling suggests that the high-altitude variant, Epas1^H, is the derived allele within the P. maniculatus subclade (S2 Table; SI Results).

Epas1 genotype is associated with physiological traits that influence oxygen homeostasis

We tested for physiological effects of allelic variation at Epas1 using deer mice captured on the summit of Mt. Evans. Epas1^H and Epas1^L alleles segregate in this population, occurring at frequencies of 0.83 and 0.17, respectively (Fig 1), which allowed us to isolate the effects of genotype on an otherwise randomized genomic background. We developed a restriction digest protocol to genotype mice in the field (see Methods), and then subjected mice of known genotype to a series of tests to characterize variation in traits that influence O₂ homeostasis under hypoxia. We used whole-body plethysmography and pulse oximetry during a step-wise hypoxia exposure [20 min at 21 (sea-level), 12 (equivalent to Mt. Evans altitude), 10, 8, and 6 kPa] to test for genotypic effects on acute respiratory and cardiovascular responses to hypoxia: breathing frequency and tidal volume, rate of O₂ consumption (VO₂), arterial O₂ saturation, and heart rate. Forty-eight hours following plethysmography and pulse oximetry, mice were exposed to deep hypoxia (6 kPa) for 2 hours and then euthanized. This duration of exposure to 6 kPa O₂ was chosen in order to strongly stimulate the HIF-induced transcriptional response, as supported by previous observations that this hypoxic treatment strongly induces the expression of Vegfa, a key HIF target, in many organisms [44]. Blood, heart, adrenal glands, lungs, and gastrocnemius muscle were sampled for transcriptomic profiling and/or for morphological and histological analysis.

We did not detect significant effects of allelic variation at Epas1 on some traits that are indicative of chronic exposure to hypoxia or that otherwise influence O₂ transport and utilization. Unlike in studies of indigenous Tibetans [12], Epas1 genotype did not affect hematocrit or hemoglobin concentration at high altitude (S3 Table, S4 Table). Similarly, we did not detect any genotypic differences in the fiber size (S8 Fig, S9 Fig) or the activities of oxidative enzymes in the gastrocnemius muscle (S5 Table, S10 Fig). There were trends toward reduced muscle capillarity in individuals that were homozygous for the lowland allele (Epas1^L/L), and reduced lactate dehydrogenase activity in the muscle of individuals that were homozygous for the highland allele (Epas1^H/H), but these differences were not statistically significant when compared to all other genotypes (S8 Fig, S9 Fig, S10 Fig).

With respect to respiratory and cardiovascular function under acute hypoxia, we also did not detect any genotypic differences in breathing (total ventilation, breathing frequency, tidal volume), VO₂, body temperature depression, or pulmonary O₂ extraction (S11 Fig, S12 Fig, S6 Table). However, Epas1 genotype did affect heart rate under ecologically relevant levels of hypoxia: resting heart rates in normoxia (21 kPa O₂) were similar across genotypes, but individuals that were homozygous for the highland allele (Epas1^H/H) maintained higher resting heart rates during hypoxia exposure compared to the other two genotypes (Epas1^L/-) (Fig 2). We detected a significant main effect of genotype on heart rate (F_2,148 = 3.8; p = 0.03), but a non-significant interaction (p>0.05), suggesting that Epas1^H/H mice generally maintained higher resting heart rates. The genotypic difference in heart rate was most pronounced at 12 kPa O₂, which approximates PO₂ at the summit of Mt. Evans (Fig 2). However, the magnitude of the increase in heart rate from normoxia to 12 kPa O₂ was greatest in Epas1^H/H mice (S13 Fig), suggesting that Epas1 genotype may have influenced the heart rate response to hypoxia at the environmental PO₂ that are realistic at the high-altitude field site. Heart rate did not differ between heterozygotes and homozygotes for the lowland allele at any PO₂ (Fig 2). We observed a steady decline in resting heart rate for all three genotypes as the level of hypoxia increased at PO₂ below 10 kPa (main effect of PO₂: F_4,148 = 19.741, P<0.001), likely as a consequence of the depression in VO₂ and body temperature under extreme hypoxia (S12 Fig).

Download:

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

Fig 2. Deer mice that were homozygous for the highland Epas1 variant exhibited higher heart rates when exposed to environmentally realistic levels of hypoxia at 4300 m altitude (12 kPa O₂).

Measurements were made using a MouseOx Plus collar. * Significant main effect of Epas1 genotype in a mixed linear model. n = 26 Epas1^H/H, n = 13 Epas1^H/L, and n = 4 Epas1^L/L variants.

https://doi.org/10.1371/journal.pgen.1008420.g002

Epas1 genotype affects the regulation of HIF target and catecholamine genes

We used an RNA-seq approach (TagSeq [45]) to test whether Epas1 genotypes differ in gene regulation in two tissues affecting heart rate under hypoxia: the adrenal gland (which affects heart rate and vasoconstrictive responses by secreting catecholamines) and the left ventricle of the heart. For each tissue we measured transcriptome-wide patterns of gene expression, and the expression of two candidate gene sets: HIF target genes and genes involved in catecholamine biosynthesis, secretion, and signaling (S7 Table). Full results of candidate gene differential expression analysis are available in the online supplement (S8 Table). Overall, this analysis revealed a significant association between Epas1 genotype and the regulation of genes in HIF- and catecholamine-related pathways in both tissues.

In the adrenal gland, we detected a subtle but significant shift toward reduced expression of candidate genes in mice possessing the high-altitude Epas1 allele (Fig 3A and 3B). Kolmogorov-Smirnov (K-S) tests revealed significant differences in the distribution of log fold-change values for each comparison of Epas1^L/L vs. Epas1^H/L and Epas1^H/H in catecholamine-associated genes (Fig 3A and 3B), but not HIF targets (p>0.05). Specifically, log-fold change values were significantly more negative for candidate genes in Epas1^H/- mice compared to the transcriptome-wide background (Fig 3A and 3B), indicating a reduced expression of these genes in mice that carry Epas1^H alleles. To test the robustness of this result, we generated a null distribution of K-S test D-statistics by drawing 1000 random gene sets that were equal in size to the candidate catecholamine gene set (n = 79). This randomization procedure demonstrates that the D-statistics calculated for both genotypic comparisons were above the 99% quantile of the null distribution (Fig 3B inset). The pattern of down-regulation of catecholamine-related genes is consistent with recent findings that, compared to low-altitude deer mice from Nebraska, high-altitude mice express lower levels of DOPA decarboxylase in the adrenal medulla and exhibit reductions in catecholamine release from adrenal chromaffin cells [46]. Moreover, HIF-2α has been shown to be a positive regulator of catecholamine synthesis in adrenal chromaffin cells of rats [21], suggesting that the high-altitude Epas1 variant leads to a direct reduction in gene expression of enzymes involved in catecholamine biosynthesis, and thereby reduces circulating catecholamine levels.

Download:

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

Fig 3. Epas1 genotype affects the regulation of HIF and catecholamine genes.

Distribution of log fold-change values between Epas1^L/L v. Epas1^H/H and Epas1^H/L for candidate genes (black) compared the the background transcriptome-wide distribution (grey) for adrenal catecholamine genes (A-B) and left ventricle HIF targets (C-D). Insets show the distribution of K-S test D-statistics for 1,000 randomly permuted datasets tested against the transcriptome-wide background. Dashed indicated the 99% quantile of the null distribution, while solid lines indicate the observed D-statistic for each comparison.

https://doi.org/10.1371/journal.pgen.1008420.g003

In a related analysis, we found that positive regulators of catecholamine processes (n = 15) were downregulated in mice possessing the high-altitude allele (mean log fold-change between Epas1^H/H and Epas1^L/L: -0.04 ± 0.17), relative to genes known to be negative regulators of catecholamines (n = 13; mean log fold-change = 0.23 ± 0.19). This trend, though not statistically significant (p>0.05) is consistent with the hypothesis that possessing the high-altitude Epas1 allele is associated with a downregulation of genes that positively influence catecholamine processes. Due to the small size of the subset of genes for which a directional influence could be established, and the lack of statistical significance, these results should be viewed with caution.

In contrast to the adrenal gland, candidate gene expression in the left ventricle was significantly higher in mice possessing Epas1^H alleles, but this effect was subtle and only significant for HIF target genes (Fig 3C and 3D). K-S tests revealed a significant shift toward more positive distribution of log fold-change values in Epas1^H/H (Fig 3C) and Epas1^H/L (Fig 3D) mice. Randomization procedures indicated that the empirical D-statistic falls outside of the 99% quantile of the null distribution for both comparisons (Fig 3D). We note that several genes in the top 10% of the distribution of log fold-change in expression (adrenomedullin [Adm], endothelin [Edn1], and atrial natriuretic peptide [Nppa]; S8 Table) are known effectors of vasodilation [47–49], suggesting that improved O₂ supply to cardiac tissue may be positively influenced by Epas1 genotype.

Importantly, none of the genes in these pathways were differentially expressed among genotypes after correcting for multiple testing in either tissue (FDR >0.05), nor were any other genes in either transcriptome. This result suggests that effects of Epas1 genotype on the regulation of any single gene are weak in these tissues. We note that the O₂ pressure (6 kPa) used to stimulate gene expression was strong, and led to uniform metabolic depression across genotypes (Fig 2) potentially masking genotypic differences in expression that may exist at less extreme levels of hypoxia. Future research will examine whether 12 kPa O₂, the level leading to the largest genotypic differences in heart rate (Fig 2), elicits greater genotypic difference in the expression of HIF and catecholamine target genes, or potentially other unknown pathways. Nevertheless, our results demonstrate that the combined effects of multiple subtle, but concerted, shifts in expression of HIF target genes and genes in catecholamine-related pathways were detectable among genotypes.

Hypoxia-related genes, including Epas1, are targets of positive selection in highland deer mice

To formally test for a history of spatially varying selection on Epas1 polymorphisms, we performed a demographically-corrected selection scan. We first estimated effective population sizes (N_e), divergence times (T), and pairwise migration rates (m) for two pairs of populations (Mt. Evans-Lincoln and Mt. Evans-Merced) by modeling the folded two-dimensional site frequency spectra (2D-SFS) derived from variation at synonymous SNPs using ∂a∂i [50]. The best-fitting demographic model (S14 Fig) allowed for variation in gene flow among loci by including two symmetrical migration rate parameters applied to proportions P and 1-P of loci, following [51], which produced significantly better fit to the data (p = 0; adjusted likelihood ratio test using the Godambe Information Matrix [52]). For Mt. Evans and Merced, we estimated that approximately 84.2% of SNPs experience a relatively high migration rate (2.07 migrants/generation), while the remaining SNPs experience a substantially lower migration rate (0.08 migrants/generations). Similarly, between Mt. Evans and Lincoln, we estimate that approximately 86.4% of SNPs experience a migration rate of ~1.75 migrants/generation while remaining SNPs have a rate of ~0.08 migrants/generation. Under these models, we estimated effective population sizes of approximately 368,000 (95% CI: 266,977–470,061) for Mt. Evans, 220,000 (95% CI: 170,878–270,767) for Lincoln, and 371,000 (95% CI: 266,871–477,310) for Merced populations (S9 Table). The inferred split time between Mt. Evans and Lincoln populations was approximately 217,000 generations ago (95% CI: 180,998–254,122 generations ago) and the inferred split between Mt. Evans and Merced was 190,000 generations ago (95% CI: 127,569–253,390 generations ago) (S9 Table).

We established a null PBS distribution by simulating 500,000 neutral SNPs (85% with the high migration rates and 15% with the low migration rates) across Mt. Evans, Merced, and Lincoln populations in msms [53] under our estimated demographic model, assuming no direct gene flow between Merced and Lincoln (S15A Fig). This approach results in a conservative null model for selection because SNPs simulated under low migration rates between populations likely mimic SNPs that experience local selection. After calculating PBS values [13] from these simulated SNPs, we used the distribution to identify outliers above the simulated 99.9^th percentile (corresponding PBS: 0.252). There were 37,169 SNPs located in 6,913 genes above the 99.9^th percentile of the simulated distribution (red dashed line in S15 Fig). The ^Thr755^Met Epas1 SNP was highly significant (S15B Fig), as were the two linked noncoding SNPs (S15B Fig). These results provide strong evidence for spatially-varying selection on Epas1 genotype and suggest that the associated phenotypic effects have fitness consequences in the wild.

Although we had a specific focus on the role of Epas1 in physiological adaptation to high altitude, our demographically-corrected exome scan also identified a number of other promising candidates for high-altitude adaptation. Among the PBS outliers relative to the simulated null, 37 GO categories and 13 KEGG pathways were significantly enriched (FDR p-value <0.05; S10 Table). Many of these enriched categories are related to known physiological mechanisms for maintaining homeostasis under hypoxia, such as “response to oxygen-containing compound” (GO:1901700), “regulation of systemic arterial blood pressure” (GO:0003073), and “circulatory system process” (GO:0003013). We focused our examination on outlier SNPs that are located within 1,247 hypoxia-related genes identified by Zhang et al. [14] (S11 Table) that represent a set of candidates compiled from “hypoxia” and “hypoxia inducible factor” keyword searches in multiple sources [14]. Of the 6,913 genes with outlier SNPs, 353 genes overlapped with the set of 1,247 hypoxia-related genes from Zhang et al. [14] (S11 Table; S12 Table; Fig 4), representing a significant enrichment of hypoxia-related genes (one-sided Fisher's exact test; FDR-corrected p-value <0.001). Epas1 was the tenth highest ranking gene (S12 Table).

Download:

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

Fig 4. Epas1 is an outlier in a demographically-controlled selection scan.

Exome-wide distribution of population branch statistic (PBS) for Mt. Evans mice based on 5,182,530 SNPs from the exome. Horizontal line (red) indicates the top 0.1% percentile PBS of 500,000 neutral SNPs simulated under a realistic demographic scenario, and SNPs in hypoxia-related genes are highlighted in orange. The top ten hypoxia-related genes are labeled, with Epas1 indicated in bold text. SNPs with a PBS value below the top 0.1% percentile have been randomly down-sampled (1:20) for ease of plotting.

https://doi.org/10.1371/journal.pgen.1008420.g004

The top three hypoxia-related genes were Collagen type I alpha 2 (Col1a2) (highest-ranking SNP (PBS: 3.33), Integrin subunit alpha 7 (Itga7) (PBS: 3.18) and Potassium calcium-activated channel subfamily M alpha 1 (Kcnma1) (PBS: 2.82). ColIa2 encodes one of the most abundant types of collagens that is a ubiquitous component of connective tissue in several organs, whereas Itga7 and Kcnma1 may both play roles in muscle structure and function. Itga7 is expressed in skeletal and cardiac muscles, where it may play a developmental role [54], and Kcnma1 encodes the pore-forming subunit of an ion channel that is integral to smooth muscle control [55]. Other high-ranking genes of note include Ceruloplasmin (Cp), which was the seventh highest ranking gene (PBS: 2.55; S12 Table) and is involved in iron transport across the cell membrane, in stimulation of erythropoiesis, and in nuclear translocation of HIF1 [56]. Further exploration of these hypoxia-related genes, as well as other targets of selection in high-altitude deer mice, is the focus of an ongoing study.

Ethics statement

We handled and sampled mice in accordance with the University of Montana's Institutional Animal Care and Use Committee (AUP 029–16), the Colorado Department of Fish and Wildlife (License Number: 17TR2168a) and the United States Forest Service (Authorization ID: CLC772). Following protocol, mice were euthanized with an isofluorane overdose followed by decapitation.

Exome design, capture, and high-throughput sequencing

To identify all annotated exons, we downloaded the Peromyscus maniculatus bairdii GFF v101 from NCBI (Accession GCF_000500354.1), and extracted all features annotated as an exon. The final set of unique, non-pseudogenized exonic regions consisted of 218,065 exons in 25,246 genes. A custom Roche NimbleGen SeqCap EZ Library kit capture a total of 226,973 regions (77,559,614 bp).

We extracted DNA from tissues of 85 deer mice (Lincoln, NE: n = 37; Mount Evans, CO: n = 48) and sheared DNA to ~300 bp using a Covaris E220 Focused Ultrasonicator. Genomic libraries for each individual were prepared using 200 ng of sheared DNA with a NEBNext UltraII kit and unique index following manufacturer’s protocols. We pooled batches of 24 indexed libraries prior to target enrichment and PCR amplification following the NimbleGen Seq Cap EZ protocol (Roche). Quality control for each capture pool included a check of size distribution and a check for enrichment of targeted regions and no enrichment of non-targeted regions using qPCR. Each capture pool of 24 individuals was sequenced with 100 bp paired-end sequencing on an Illumina HiSeq 4000. We extracted and quantified DNA samples for the California deer mice (Merced, CA; n = 15) at the Museum of Vertebrate Zoology, UC Berkeley, before shearing one μg of genomic DNA to less than 500 bp with a Biorupter (Diagenode). We prepared barcoded Illumina sequencing libraries using the Meyer and Kircher [63] protocol, then amplified libraries with Phusion High-Fidelity DNA Polymerase (Thermo Scientific) for 6–8 cycles during the indexing PCR. Exome enrichment was conducted with a custom capture design from the SeqCap EZ Developer Libary (Nimblegen) that was almost identical to that used in the 85 non-CA samples. Captures were quantified, pooled proportionally to the amount of DNA in each, and sequenced using 100bp pair-end sequencing on an Illumina HiSeq4000.

Data pre-processing and variant discovery on all samples followed the recommendations of the Broad Institute GATK v3.7-0-gcfedb67 Best Practices pipeline. We trimmed reads of adapter sequences and for a minimum base quality of 20 using fastq_illumina_filter 0.1 (https://mcbl.readthedocs.io/en/latest/mcbl-tutorials-PF-clean.html) and trim_galore 0.3.1 (http://www.bioinformatics.babraham.ac.uk/projects/trim_galore/). We used bwa mem [64] to align and map forward and reverse reads to the Peromyscus maniculatus baiardii genome. We removed duplicates using samtools rmdup [65], then added read group information using picard. (http://picard.sourceforge.net). To generate a set of “known” variant sites for GATK Base Quality Score Recalibration (BQSR), we genotyped each individual using samtools mpileup (-q 30 -Q 30) and bcftools call, then filtered genotypes to have a minimum depth of coverage (minDP) of 10 and minimum genotype quality (minGQ) of 30, and only used those variants observed in at least two individuals. The resulting set of variant positions was used with the -knownSites flag during GATK BQSR. A subsequent round of BQSR was completed and convergence of quality scores was verified using GATK AnalyzeCovariates. To genotype each sample, we used GATK HaplotypeCaller with the ‘--emitRefConfidence' flag, then called variants GATK GenotypeGVCFs. We combined GVCFs and filtered them to remove SNPs with a quality of depth <2.0, a FS > 60, mapping quality < 40, mapping quality rank sum < -12.5, and read position rank sum < -8.0. We implemented all processing steps in GATK using the ‘--interval' flag, a bed file of capture regions, and a ‘--interval_padding’ of 200 bp. These processing steps resulted in a total of 106,883,914 sites among all individuals.

After assessing the quality of filtered reads using the vcftools package[66], we further filtered variants so that a site was called in at least 50% of individuals, was bi-allelic, and each site had a minDP of 5 and minGQ of 20. We proceeded with a set of 5,183,434 high-quality bi-allelic variants, with a mean depth of coverage of 18.10±6.38 X (S1 Fig).

Population genetic structure

We assessed population genetic structure of Mount Evans, Lincoln, and Merced mice using principal components analysis within PLINK [40] and Admixture [41]. Prior to running the analyses, we pruned the set of variants to only sites with no missing data, and not linked (using the “--indep-pairwise 50 5 0.5” option within PLINK). The final set of variants for these analyses consisted of 296,196 bi-allelic sites.

Allele frequency variation in HIF genes

In order to examine altitudinal patterns of allele frequency variation of the genes encoding HIF-1α, HIF-2α, and HIF-3α, we calculated pairwise F_ST using vcftools, then the Population Branch Statistic (PBS) [13] for each of the 5,183,434 bi-allelic variants. Using the ‘ecdf’ function in R, we used the empirical distribution of PBS values and set a threshold of 99.9% for significance (corresponding PBS: 0.886). We used Ensembl’s Variant Effect Predictor [67] with the P. maniculatus reference genome annotation data set to identify SNPs located in HIF genes. For the three outlier SNPs located in Epas1, we calculated pairwise linkage between each SNP as the squared correlation coefficient, r², using the ‘--geno-r2’ function within vcftools [66],

Geographic and phylogenetic survey

To fully assess the geographic and phylogenetic extent of variation in Epas1, we genotyped 266 P. maniculatus samples from across the western US (S1 Table), plus samples of P. keeni, P. melanotis, P. hylocytes, P. attwateri, P. melanophrys, P. eremicus, P. polionotus, P. leucopus, Reithrodontomys montanus, and Phyllotis xanthopygus (S2 Table). We obtained tissue samples from our existing freezer collections or museums (Museum of Southwestern Biology at the University of New Mexico, or Museum of Comparative Zoology, Harvard University). From each sample, we extracted DNA then PCR amplified Epas1 with custom exonic primers (“epas1_snp_L” and “epas1_snp_R”; S13 Table) designed from the P. maniculatus bairdii genome under the following conditions: 94°C for 2 mins; 30 cycles of 94°C for 45 sec, 58°C for 1 min, 72°C for 1 min; then 72°C for 10 mins. To improve amplification specificity for P. maniculatus samples, we used modified primers and PCR conditions (“epas1_set1_F” and “epas1_set1_R”; S13 Table): 94°C for 2 mins; 35 cycles of 94°C for 30 sec, 62°C for 30 sec, 68°C for 1 min; then 68°C for 10 mins. Technicians at Genewiz (South Plainfield, NJ) cleaned amplified products and sequenced them in both directions. We called genotypes after aligning sequences to the reference sequence using Geneious 8.1.8. We calculated population allele frequencies within each sampling locality, and obtained elevation for each locality from GPS data recorded upon sampling, or from Google Maps. We mapped these allele frequency data on a map of elevation (data downloaded from www.worldclim.org) using the ‘maps’ package in R. Finally, we placed Epas1 genotype for each sequenced species on a Peromyscus phylogeny constructed previously [68–70].

Cline analysis for Epas1 and Hemoglobin genes

We tested whether the Epas1 allele frequencies follow a clinal pattern using the R package HZAR [71]. In HZAR, genetic data are fit to equilibrium cline models using the Metropolis–Hastings Markov chain Monte Carlo (MCMC) algorithm, and parameters such as the cline center (c) and width (w) are estimated. c and w characterize the location within the transect where the variable changes most rapidly, and the values of these parameters can be estimated within HZAR by 15 models that differently estimate the exponential decay on either side of the cline center, as well as the minimum or maximum frequencies. We used as input the Epas1 allele frequency data and elevation for populations of deer mice sampled across the Rocky Mountain to Great Plains elevational cline, and used a burn-in of 10000. We compared the Epas1 cline to that for previously published hemoglobin haplotype frequencies in deer mice [43].

Physiological effects of allelic variation at Epas1

To test for physiological effects of allelic variation at Epas1, we captured deer mice from a single interbreeding population in which the high-altitude allele is segregating. We collected adult deer mice from the summit of Mt. Evans (Clear Creek Co., Colorado, USA; 39°35’18” N, 105°38’38” W; 4,350 m above sea level; PO₂ ~ 95.6 mm Hg) in August 2016 and 2017 and screened the Epas1 allele (a C/T polymorphism at nucleotide position 2264, hereafter ^C2264^T). From each individual, we sampled an ear clip sample, extracted DNA, then genotyped Epas1 ^C2264^T using a custom restriction enzyme digest assay. Briefly, we PCR amplified Epas1 with custom exonic primers (S13 Table) designed from the P. maniculatus bairdii genome and amplified under the PCR conditions specified above. For all amplified PCR products, we cut Epas1 at ^C2264^T by incubating the PCR product with the BsaHI restriction enzyme at 37°C for 1 hour followed by a heat denaturation for 20 mins. We called Epas1 genotypes via gel electrophoresis (T ~675 bp; C ~300 bp;), then subsequently confirmed field genotypes with Sanger sequencing at Genewiz.

Acute hypoxia responses with pulse oximetry

At the University of Denver Mt. Evans field station (3230 m a.s.l.; ~15kPa O₂), we screened these Mt. Evans mice with alternative Epas1 genotypes for a suite of physiological responses involved in O₂ transport and utilization and/or known to be influenced by HIFs. We measured hypoxia responses in mice (26 Epas1^H/H, 13 Epas1^H/L, and 4 Epas1^L/L) using previously described barometric plethysmography, respirometry, and pulse oximetry techniques [34,72]. We placed each mouse in a whole-body plethysmograph (chamber volume: 530 ml) that was supplied with hyperoxic air, mixed to simulate the partial pressure of O₂ (PO₂) at sea level (21 kPa O₂, balance N₂), at a rate of 600 ml min^-1. We gave mice 20–60 min to adjust to the chamber and stabilize their breathing and metabolism. We recorded measurements for an additional 20 min at 21 kPa O₂, then exposed mice to 20 min stepwise reductions in inspired PO₂ of 12, 10, 8, and 6 kPa. We set the incurrent gas composition by mixing dry compressed gases using precision flow meters and a mass flow controller, such that the desired PO₂ was delivered to the chamber at a constant rate of 600 ml min^-1. At the end of the experiment, we measured body temperature (T_b) using a mouse rectal probe. We also measured T_b exactly 24 h later to determine resting T_b (this was used as a proxy for the resting T_b at the start of the experiment, which was not measured to prevent stress to the animal).

We determined breathing and O₂ consumption rate (VO₂) during the last 10 min at each PO₂ by subsampling incurrent and excurrent airflows at 200 ml min^-1. For incurrent and excurrent air, we measured water vapor (RH-300, Sable Systems) using a thin-film capacitive water vapor analyzer, then dried air with pre-baked drierite, and measured continuously for O₂ and CO₂ fraction using a galvanic fuel cell O₂ analyzer and infrared CO₂ analyzer (FOXBOX, Sable Systems). We used these data to calculate VO₂ and CO₂ production rate (VCO₂), expressed at standard temperature and pressure (STP), using appropriate equations for dry air as described by Lighton [73]. We measured breathing frequency and tidal volume from changes in flow across a pneumotachograph in the plethysmograph wall, detected using a differential pressure transducer. We calculated tidal volume using established equations [75,76] and assuming a constant rate of decline in T_b with declining PO₂, which we have previously shown results in similar tidal volumes to those calculated using direct T_b measurements at each PO₂ [72], and is expressed at STP. We calculated the following parameters: total ventilation (the product of breathing frequency and tidal volume), ventilatory equivalent for O₂ (total ventilation divided by VO₂), and percent pulmonary O₂ extraction (VO₂ divided by the product of total ventilation and inspired O₂ fraction). We measured SaO₂ and heart rate using the MouseOx Plus pulse oximeter collar sensors and associated software (Starr Life Sciences, Oakmont, PA, USA). Use of the collars was enabled by removing a small amount of fur from the skin around the neck one day before experiments.

We tested for effects of Epas1 genotype on cold-induced VO₂ max (thermogenic capacity), an ecologically relevant measure of whole-organism aerobic performance for which there is an evolved difference between lowland and highland deer mice [24]. To do this we measured maximum rates of oxygen consumption in a hypoxic, heliox atmosphere (21% oxygen, 79% helium) using open-flow respirometry. All trials were conducted at the summit of Mt. Evans. The use of heliox ensures that VO₂ max can be measured without risking cold injury, since rates of heat loss in heliox are several times greater than that of ambient air [74]. For each trial, we equilibrated heliox gas mixtures with atmospheric pressure of the Mt. Evans summit (12 kPa). Mass flow controllers helped pump the heliox mixture into copper coils inside a temperature control chamber. The cooled gas was pumped into an animal chamber and an empty baseline chamber at a rate of ~750 ml min^-1. Excurrent air from the animal and baseline chambers was sampled at a rate of ~130 ml min^-1, dried with magnesium perchlorate and scrubbed of CO₂, redried with drierite, and passed through an oxygen analyzer. We defined thermogenic capacity as the maximum VO₂ averaged over a continuous 5-min period. We tested for the influence of genotype on thermogenic capacity using an analysis of covariance (ANCOVA) with body mass as covariate.

Tissue and organ sampling and phenotyping

Mice recovered for 2–3 days after the hypoxia response experiments described above. We exposed recovered mice to 2 hours of deep hypoxia (6 kPa O₂) and euthanized them with an isofluorane overdose followed by decapitation. We collected blood samples for hematocrit (in heparinized capillary tubes, spun for 5 minutes) and hemoglobin content (Hemocue, Sweden). We dissected the heart, then isolated and weighed the ventricles before freezing them separately in liquid N₂. We determined lung volume by volumetry [75]. We weighed and froze in liquid N₂ one gastrocnemius muscle, then coated the other in embedding medium and froze it in liquid N₂-cooled isopentane for histology. We froze other organs directly in liquid N₂.

We measured muscle capillarity using histological methods in a subset of the mice (17 Epas1^H/H, 13 Epas1^H/L, and 4 Epas1^L/L) chosen to ensure a balanced sex distribution (11 Epas1^H/H males, 6 Epas1^H/H females, 11 Epas1^L/- males, 6 Epas1^L/- females) and body mass (21.49 ± 4.05 g for Epas1^H/H, 21.76 ± 5.43 g for Epas1^H/L, 21.01 ± 4.12 g for Epas1^L/L; means ± SEM) between the three groups. We prepared full transverse sections of the gastrocnemius muscle as previously described [32,76]. Briefly, after cutting 10 μm tissue sections transverse to the muscle fiber length using a cryostat, we identified capillaries by staining samples for alkaline phosphatase activity following previous studies [32,76]. We used bright-field microscopy to systematically collect images from across the entire gastrocnemius, and used ImageJ software [77] to count the number of capillaries and muscle fiber in each image and measured capillary density, average number of capillaries per muscle fiber, and average transverse area of muscle fibers. We used NIS-Elements D Imaging Software (v. 4.30, Nikon Instruments) to measure the number, perimeter, and capillary surface densities of individual capillaries within each image. We determined a sufficient number of images to analyze to account for heterogeneity across the gastrocnemius, determined by the number of replicates necessary to yield a stable mean value, following ref. [32]. This required analysis of roughly half of the entire section, with images spread evenly across the section, which was found to be more than sufficient to accurately represent average values across the entire muscle. For all histological measurements, the observer was blind to genotype during analysis.

We used the remaining gastrocnemius muscle tissue in metabolic enzyme assays that we have previously described [76]. After removing embedding medium from the muscle tissues we powdered samples under liquid N₂ and homogenized them in ice-cold homogenization buffer [76]. We centrifuged homogenates at 1000g for 1 min at 4°C, discarded the pellet, and stored the homogenate on ice until assay. We assayed activities of cytochrome c oxidase (COX), citrate synthase (CS), and lactate dehydrogenase (LDH) in triplicate at 37°C using a 96-well microplate reader. Assay conditions in mM were as follows: COX, 100 KH₂PO₄, 0.2 reduced cytochrome c*, pH 8.0; CS, 100 KH₂PO₄, 0.5 oxaloacetate*, 0.15 acetyl-coA, 0.15 5,5′-dithiobis-2-nitrobenzoic acid (DTNB), pH 8.0; LDH, 100 KH₂PO₄, 0.15 NADH, 2.5 pyruvate*, pH 7.2. We determined maximal activities by measuring the change in absorbance over time at 550 nm for COX (ε = 28.5 mM^-1 cm^-1), 412 nm for CS (ε = 14.15 mM^-1 cm^-1), and 340 nm for LDH (ε = 6.22 mM^-1 cm^-1), and subtracting the background rate from the rates measured in the presence of all substrates.

Statistical analyses of physiological effects of allelic variation at Epas1

To assess the influence of Epas1 genotype on physiology and tissue/organ phenotypes, we used linear mixed effect models and included body mass, genotype, and acute PO₂ (when appropriate) as fixed effects. We initially included year (2016 or 2017) and sex as random effects, but removed them from all models because they were never found to be significant (P>0.25). We removed body mass from models in which its effect was not significant for variables that we did not have any a priori expectation of allometric scaling (heart rate, SaO₂, hematology, and gastrocneumius muscle capillarity and enzyme activities). We conducted Holm-Sidak pairwise post-tests on significant models, and used R (v. 3.4.3) and the lme4 package for all statistical analysis, with a significance level of 0.05. We report VO₂, total ventilation, and tidal volume relative to body mass to enable comparison to the literature, but we used the absolute data (i.e., not expressed relative to body mass) for statistical analyses as described above.

Transcriptomic analysis of differential gene expression

We used high throughput sequencing to test for effects of Epas1 genotype on gene expression in adrenal gland (8 Epas1^H/H; 8 Epas1^H/L; 3 Epas1^L/L) and heart tissue (7 Epas1^H/H; 9 Epas1^H/L; 3 Epas1^L/L). We chose the adrenal gland because of its role in stimulating heart rate via catecholamine release. We assayed gene expression using TagSeq, a 3’ tag-based sequencing following ref. [45]. We extracted RNA from 25 mg of tissue using TRI Reagent (Sigma-Aldrich), then assessed RNA quality using TapeStation (Agilent Technologies; RIN > 7). The Genome Sequencing and Analysis Facility at the University of Texas at Austin prepared TagSeq libraries, which were sequenced using Illumina HiSeq 2500. Sequencing generated an average of 4.6M reads per individual. We processed raw reads following Lohman et al. [45] and mapped them to the P. maniculatus genome using bwa [64]. We used featureCounts [78] to generate a table of transcript abundances. Since genes with low read counts are subject to increased measurement error [79], we excluded those with less than an average of 10 normalized reads per individual using the filterByExpr function in edgeR. We retained a total of 12,237 and 10,509 genes after filtering for adrenal and heart transcriptomes, respectively.

We used two complementary approaches to compare levels of transcript abundance among Epas1 genotypes: (1) A whole-transcriptome differential expression analysis was conducted to identify genes that were differentially expressed in each tissue. (2) We performed candidate differential expression analysis on two a priori gene sets aimed at testing whether genes related to the HIF cascade and/or catecholamine synthesis/transport exhibited concerted changes in gene expression among alternative genotypes. We conducted the whole-transcriptome differential expression analysis in edgeR [80]. The function calcNormFactors was used to normalize read counts among all libraries, after model dispersion was estimated for each transcript separately using the function estimateDisp [81]. We tested for differences in transcript abundance by first fitting a quasi-likelihood negative binomial generalized linear model to raw count data (glmQLFit function), which included a single main effect of genotype (Epas1^L/L was used as a reference for comparing against Epas1^H/L and Epas1^H/H). P-values were calculated using a quasi-likelihood F test using the glmQLFTest function. We controlled for multiple testing by enforcing a genome-wide false discovery rate correction of 0.05 [82]. We identified candidate HIF target genes from the literature [83,84] and Kyoto Encyclopedia of Genes and Genomes database [85] as those with known function in HIF signaling, and those that have an unknown function but contain HIF binding sites [84]. We ascertained catecholamine-related genes based on annotation in the Gene Ontology (GO) database in AmiGO (amigo.geneontology.org). A total of 277 HIF targets and 149 catecholamine related genes (S7 Table) were identified.

To determine whether there were concerted shifts in gene expression for the candidate gene sets among genotypes, we calculated log fold-change in expression between Epas1^L/L mice and mice heterozygous and homozygous for the high-altitude allele in edgeR. The distribution of fold-change values between candidate gene sets and the transcriptome-wide background (candidate genes excluded) was compared using Kolmogorov-Smirnov (K-S) tests using the function ks.test in R. We then conducted a randomization procedure to test whether the observed D-statistic for K-S tests was greater than a null distribution; the null distribution of D-statistic values was produced by 1000 random draws of gene sets that were of equal size to candidate sets (adrenal: HIF: n = 207; catecholamine: n = 79; left ventricle: HIF: n = 207; catecholamine: n = 55) and comparing those to the transcriptome-wide background.

Demographic modeling and null PBS distribution

We estimated the demographic history of highland and lowland deer mice using synonymous SNPs in ∂a∂i [50]. We filtered SNPs in Hardy-Weinberg equilibrium (p<0.001) and excluded sites with >25% missing data per population using vcftools, resulting in 287,336 SNPs to generate a folded site-frequency spectrum. We then estimated effective population sizes (N_e), divergence times (T), and pairwise migration rates (m) between highland deer mice from Mt. Evans, CO (n = 48) and deer mice from Lincoln, NE (n = 37) and Merced, CA (n = 15). We assumed that any migration between Lincoln and Merced populations would occur indirectly through the central Mt. Evans population and thus we did not perform a pairwise demographic analysis for Lincoln and Merced. For our pairwise population comparisons (Mt. Evans-Lincoln and Mt. Evans-Merced), we calculated maximum-likelihood (ML) parameters for demographic models with and without a single symmetrical migration parameter and with an effective population size parameter (u; proportional change in N_e relative to the ancestral population immediately following the split). We observed that maximum likelihood parameters under models of no migration or a single symmetrical migration rate strongly underestimated the relative abundance of highly differentiated SNPS, resulting in poor fit to the empirical 2D-SFS (S14 Fig). We also tested a model which included heterogeneous migration rates among loci in the genome. Here, we included two symmetrical migration rates, one for proportion P of SNPs and one for proportion 1-P of SNPs, where we also estimate the P parameter. For each demographic model, we performed 25 independent runs with starting parameter values sampled randomly from a uniform prior distribution (0<2N_em<10; 0.01<2N_et<10; 0.1<u<10; 0.5<P<1.0). We selected the optimal demographic model based on an adjusted likelihood ratio test using the Godambe Information Matrix [52]. We estimated 95% confidence intervals for population size, divergence time, and migration rate parameters using the Godambe Information Matrix with 100 replicate bootstrap data sets consisting of randomly sampled SNPs spaced at least 10 kb apart (14250 SNPs for each data set). We calibrated theta estimates based on the ratio of all callable sites to SNPs under the same filtering regime and assuming a mutation rate of 5.4x10^-9 per base per generation (house mouse; ref. [86]).

To establish a null PBS distribution, we simulated 500,000 neutral SNPs across three populations in msms [53] under our estimated demographic model. Given our optimal models included two symmetrical migration rate parameters applied to different sets of SNPs we simulated proportion P of SNPs under high migration rates (85%) and proportion 1-P of SNPs under low migration rates (15%) and combined the two simulated data sets. We used msstats (https://github.com/molpopgen/msstats) to obtain F_ST values for SNPs between each population and calculated PBS based on the equation in Yi et al. [13].

Demographically corrected exome scan with the Population branch statistic

We used the simulated distribution of PBS values, and set a significance threshold of 99.9% (corresponding PBS_sim: 0.199). We focused our examination on outlier SNPs that are located within 1,247 hypoxia-related genes from Zhang et al. [14] (S11 Table). The genes from Zhang et al. represent a set of candidates compiled from “hypoxia” and “hypoxia inducible factor” keyword searches in multiple sources, including the UCSC Genome Browser, Ensembl, NCBI, UniProt, and RefSeq. For each P. maniculatus gene containing outlier SNPs, we found the corresponding Mus musculus gene, then used gProfiler [87] to identify enriched gene ontology categories above a false discovery rate corrected significance of 0.05, using strong hierarchical filtering.