Adaptive Shifts in Gene Regulation Underlie a Developmental Delay in Thermogenesis in High-Altitude Deer Mice

Abstract

Aerobic performance is tied to fitness as it influences an animal’s ability to find food, escape predators, or survive extreme conditions. At high altitude, where low O₂ availability and persistent cold prevail, maximum metabolic heat production (thermogenesis) is an aerobic performance trait that is closely linked to survival. Understanding how thermogenesis evolves to enhance survival at high altitude will yield insight into the links between physiology, performance, and fitness. Recent work in deer mice (Peromyscus maniculatus) has shown that adult mice native to high altitude have higher thermogenic capacities under hypoxia compared with lowland conspecifics, but that developing high-altitude pups delay the onset of thermogenesis. This finding suggests that natural selection on thermogenic capacity varies across life stages. To determine the mechanistic cause of this ontogenetic delay, we analyzed the transcriptomes of thermoeffector organs—brown adipose tissue and skeletal muscle—in developing deer mice native to low and high altitude. We demonstrate that the developmental delay in thermogenesis is associated with adaptive shifts in the expression of genes involved in nervous system development, fuel/O₂ supply, and oxidative metabolism pathways. Our results demonstrate that selection has modified the developmental trajectory of the thermoregulatory system at high altitude and has done so by acting on the regulatory systems that control the maturation of thermoeffector tissues. We suggest that the cold and hypoxic conditions of high altitude force a resource allocation tradeoff, whereby limited energy is allocated to developmental processes such as growth, versus active thermogenesis, during early development.

Introduction

Fitness in the wild is determined by suites of interacting traits that influence variation in whole-organism performance. Performance in turn determines an organism’s ability to conduct ecologically relevant tasks such as avoiding predators, competing for resources, and surviving extreme events (Huey and Stevenson 1979; Garland and Losos 1994; Irschick and Garland 2001; Irschick et al. 2008; Campbell-Staton et al. 2017). Many of these ecologically relevant tasks that impinge on whole-organism performance are ultimately dependent on the capacity for aerobic metabolism, and thus, the ability to transport and utilize O₂. As such, the development of the physiological machinery needed to deliver O₂ and metabolize fuels is closely tied to both postnatal juvenile survival and the future reproductive success as adults. The timing of key developmental events and responses to early environmental exposures should therefore influence the evolution of whole-organism performance (Lailvaux and Husak 2014). Studies that seek to understand how performance evolves in the context of development are rare but are needed to determine the causal connections between phenotype and fitness in the wild.

Extreme environments, such as high-altitude habitats >3,000 m above sea level (a.s.l.; Bouverot 1985), are windows into the mechanisms that shape adaptive variation in performance, since the selection pressures are few in number and strong in magnitude (Garland and Carter 1994). The agents of selection at high altitude—cold temperature and unavoidable reductions in available O₂ (hypobaric hypoxia)—have led to clear examples of local adaptation in animals that live there permanently (e.g., Beall 2007; Storz et al. 2010; Simonson et al. 2012). A considerable amount of recent work on high-altitude adaptation has been conducted on the North American deer mouse (Peromyscus maniculatus) owing to its broad altitudinal distribution, ranging from sea level to >4,000 m a.s.l. (King 1968). Recent comparative physiological work demonstrates that adult high-altitude deer mice from the Colorado Rocky Mountains have consistently higher aerobic performance capabilities than conspecifics in the Great Plains: High-altitude deer mice are capable of greater aerobic performance under hypoxia and in response to extreme cold (referred to as thermogenic capacity) measured in the wild and in the laboratory (Cheviron et al. 2012, 2013, 2014; Lui et al. 2015; Lau et al. 2017; Tate et al. 2017). Capture–mark–recapture data confirm that higher thermogenic capacities are beneficial for survival at high altitude (Hayes and O’Connor 1999). These data suggest that enhanced thermogenic capacity provides an important performance benefit in high-altitude deer mice that is strongly tied to fitness during adulthood.

Recent work has uncovered the factors that shape adaptive variation in thermogenic performance to illuminate links between physiology and fitness. This research, conducted in adult deer mice, suggests that improved thermogenic performance is related to alterations to O₂ transport and utilization, including more effective breathing patterns (Ivy and Scott 2017), higher blood-O₂ affinity (Snyder 1981; Snyder et al. 1982; Chappell and Snyder 1984; Storz 2007, 2016) and circulation (Tate et al. 2017), an improved capacity to oxidize lipids as fuel (Cheviron et al. 2012, 2014), and a greater oxidative capacity in skeletal muscles and their mitochondria (Lui et al. 2015; Scott et al. 2015; Mahalingam et al. 2017). Despite the importance of these changes in adults, it is not clear how high-altitude animals have evolved to survive the unique challenges of cold and hypoxic stress during development. Early postnatal development, however, is critically linked to fitness: Altricial rodent pups are small, immobile, and reliant on limited energy supplied through maternal care, and pup mortality can be extremely high in the wild (Hill 1983). Moreover, mouse pups are not born with the ability to generate heat (Pembrey 1895). Independent thermogenic abilities develop as a result of the maturation of two thermoeffector organs, brown adipose tissue (BAT) and skeletal muscle, which permits nonshivering thermogenesis (NST) and shivering thermogenesis, respectively (Lagerspetz 1966; Arjamaa and Lagerspetz 1979). Accordingly, if selection pressures on thermogenic performance are constant across life stages, then thermogenesis should develop faster in high-altitude pups to support an improved adult performance.

However, Robertson et al. (2019) have recently demonstrated that NST is delayed by ∼2 days in high-altitude deer mouse pups compared with lowland conspecifics and a closely related, but strictly lowland species, Peromyscus leucopus. This delay in the onset of NST is further associated with a delay in the onset of shivering thermogenesis until the normal date of weaning (Robertson and McClelland 2019; fig. 1). The authors suggest that the observed ontogenetic delay may be an evolutionary response to limitations in energy production or allocation at high altitude (e.g., as a result of low O₂), which forces a tradeoff between active thermogenesis and other developmental functions such as growth. Consistent with this interpretation, growth rates under common garden conditions are identical between low- and high-altitude pups, despite the fact that high-altitude mothers produce larger litters (Robertson et al. 2019).

Fig. 1.

The development of thermogenesis is delayed in highland (blue, circles) relative to lowland (orange, triangles) deer mice. (A) Cold (24 °C)-induced O₂ consumption rate (VO₂) across postnatal day p2–p10. Values constitute the difference between cold-induced and normothermic VO₂, which controls for the stress of separating pups from their mothers at a young age. Values crossing zero indicate that pups are able to generate heat independently. (B) Thermogenic capacity (VO₂ max during exposure to 5 °C in heliox) under normoxia is significantly lower in highlanders than lowlanders at p14. Values represent mean mass-corrected VO₂. *P < 0.05 and **P < 0.01. Data from Robertson et al. (2019) (A) and Robertson and McClelland (2019) (B).

We tested the hypothesis that the ontogenetic delay in thermogenesis is driven by gene regulatory changes that delay the development of the primary thermoeffector organs. To do this, we compared BAT and skeletal muscle transcriptomes in deer mice native with low and high altitudes across the first 27 days of life. We associated variation in transcript abundances to variation in thermogenic capabilities using a gene coexpression network approach that has been used to link gene regulation to physiological function in variety of evolutionary contexts (DeBiasse and Kelly 2016; Velotta et al. 2016, 2018; Campbell-Staton et al. 2018). We show that the delay in thermogenesis in high-altitude mice is associated with a delay in the expression of gene networks that function in nervous system control of BAT, fuel/O₂ supply to BAT, as well as aerobic metabolism and mitochondrial biogenesis in skeletal muscle. Finally, using a combination of phenotypic divergence and population-genetic approaches, we provide evidence that the ontogenetic delays in thermogenesis and their associated regulatory changes are adaptive at high altitude. By combining these approaches, we demonstrate, for the first time, that selection has altered the developmental trajectories of a thermoregulatory system by acting on the regulatory control of thermoeffector organ phenotype.

Results

The Ontogeny of Thermogenesis

We reanalyzed cold-induced metabolic rate (rates of O₂ consumption, VO₂) from Robertson et al. (2019) and Robertson and McClelland (2019) to assess the ontogeny of thermogenesis in deer mice native to lowland (320 m a.s.l.) and highland (4,350 m a.s.l.) habitats. VO₂ of pups postnatal age p0–p10 was measured during exposure to a mild cold stressor (24 °C) and was calculated as the difference in cold-induced VO₂ compared with that of normothermic (30 °C) littermates (see Materials and Methods). Cold-induced VO₂ in these young pups was ∼0 until p6 for highlanders and lowlanders (fig. 1A). After p6, cold-induced VO₂ increased in both populations but did so at a slower rate in highlanders (fig. 1). Statistically significant population differences in VO₂ were detected at p4 and p10 (fig. 1). We measured thermogenic capacity in older pups (age p14, p21, and p27) as VO₂max during exposure to cold (−5 °C) in a heliox air mixture. Highland deer mice exhibited a significantly reduced thermogenic capacity at p14 compared with lowlanders (∼6 ml O₂ per gram per hour, on average), whereas by p21 thermogenic capacities between populations were nearly equivalent (fig. 1B). Although there is a trend toward highlanders surpassing lowlanders by p27, this was not statistically significant (P = 0.2).

Transcriptional Correlates of Thermogenesis

We sequenced the transcriptomes of two thermoeffector organs, BAT and skeletal muscle, in order to assess regulatory mechanisms underlying evolutionary change in the ontogeny of thermogenesis at high altitude. Principal components analysis (PCA) of transcriptome-wide patterns of gene expression revealed separation between lowlanders and highlanders, and across ontogeny, in BAT and skeletal muscle (fig. 2). For BAT, highlanders and lowlanders were distinguishable along PC1, which explained 18% of total variation in expression. By contrast, PC2, which explained 16% of the variance, distinguished postnatal age (fig. 2A). We observed similar separations in skeletal muscle samples, although in this case populations were distinguishable along PC1 (22% of variance), and postnatal age distinguishable along PC2 (16% of variance). Ninety-five percent confidence ellipses were drawn around each population demonstrating that populations were entirely (skeletal muscle, fig. 2B) or almost entirely (BAT, fig. 2A) nonoverlapping in PC1 and PC2 space.

Fig. 2.

 PCA of transcriptome expression data for thermoeffector organs (A) BAT and (B) skeletal muscle in lowland (orange points) and highland (blue points) deer mice. Symbols represents ages from postnatal day p0–p10 for BAT, and p14–p27 in skeletal muscle. Populations separated along PC1 in BAT data and PC2 in skeletal muscle data. Gray circles are 95% confidence ellipses drawn around each population to highlight separation.

Weighted gene coexpression network analysis (WGCNA) of BAT samples across p0–p10 yielded 38 modules ranging in size from 61 to 1,073 genes (see supplementary table S2, Supplementary Material online, for a full list of module assignments). A total of 1,588 out of 11,192 genes could not be assigned to any module (module B0). We found ten modules exhibiting statistically significant associations between module eigengene and cold-induced VO₂, after correction for multiple testing (table 1). Of the ten VO₂-associated modules, six exhibited significant population effects on module eigengene (table 1). Two of these (modules B11 and B35) were not significantly enriched for Gene Ontology (GO) terms or were enriched for terms unrelated to thermogenesis (e.g., “drug metabolism”; see supplementary table S4, Supplementary Material online, for the full list of enriched terms). Modules B4 and B12 (table 1) were both significantly enriched for genes that encode ribosomal proteins (supplementary table S4, Supplementary Material online), which may indicate differences in protein translation between populations. No significant interaction effects were detected for any VO₂-associated module (table 1).

Table 1.

BAT Modules Significantly Correlated with Cold-Induced VO₂.

Module	Module Size	Population		Age		Pop × Age		VO2 Correlation
		F 1,30	P	F 1,30	P	F 1,30	P	r	P
B2	1,022	7.4	0.41	3.2	1.0	5.36	1.0	−0.49	0.02
B3	773	27.0	<0.001	20.8	<0.001	1.67	1.0	0.67	<0.001
B4	252	18.6	0.01	38.1	<0.001	0.43	1.0	−0.56	0.004
B11	99	162.1	<0.001	26.8	<0.001	0.04	1.0	−0.47	0.02
B12	605	21.5	<0.001	128.2	<0.001	4.90	1.0	−0.81	<0.001
B16	149	2.7	1.0	25.9	<0.001	0.06	1.0	−0.55	0.01
B32	267	0.1	1.0	32.9	<0.001	0.36	1.0	−0.52	0.01
B33	1,073	0.7	1.0	53.9	<0.001	0.85	1.0	0.64	<0.001
B35	74	37.3	<0.001	0.9	1.0	0.02	1.0	−0.45	0.03
B36	708	26.5	<0.001	51.5	<0.001	1.52	1.0	0.58	<0.001

Note.—Full association test results are presented in supplementary table S3, Supplementary Material online. Effect of population (highland vs. lowland), age (postnatal day p0–p10), or their interaction from ANOVA models on rank-transformed module expression values (module eigengene). Final two columns present results of Pearson correlation between module eigengene and cold-induced VO₂. P values from association tests and ANOVAs were corrected for multiple testing using the false-discovery rate method. Significant P-values are italicized.

Modules B3 and B36, by contrast, were positively associated with cold-induced VO₂ (fig. 3A and B), expressed at significantly lower levels in highlanders compared to lowlanders (fig. 3B and C), and exhibited significant enrichment for functions that reflect a delay in the development of BAT in highlanders (fig. 3E and F). Module B3 was enriched for functional terms related to fatty acid metabolism in the mitochondria, including the GO Biological Process term “fatty acid metabolic process,” the GO Cellular Component term “mitochondron,” the KEGG pathway “fatty acid degradation” (fig. 3E), and Reactome and WikiPathways terms “fatty acid metabolism” and “fatty acid beta oxidation” (supplementary table S4, Supplementary Material online). Module B36 was enriched for terms related to the vascularization of BAT and the development of its neural circuitry (fig. 3F): vascularization terms include the GO Biological Processes “blood vessel development” and “vasculature development,” as well as the KEGG pathway “vasculature endothelial growth factor (VEGF) signaling,” which is the major pathway regulating new blood vessel growth. Indeed, Vegfa exhibits significantly lowered expression in highland compared with lowland mice (fig. 3F inset), though it was not assigned to module B36 (supplementary table S2, Supplementary Material online). Finally, we detected enrichment of the KEGG pathway “axon guidance,” which represents a key stage in the development of neuronal networks.

Fig. 3.

Two BAT modules are associated with variation in thermogenesis in p0–p10 deer mouse pups. (A, B) The significant positive association between module eigengene and cold-induced VO₂ (measure of thermogenesis). Eigengene values of modules B3 (C) and B36 (D) are significantly affected by age, and significantly differentiated between lowland (orange, triangles) and highland (blue, circles) populations. We detected significant enrichment of KEGG pathways related to metabolism and fatty acid degradation in B3 (E), and VEGF signaling and axon guidance in B36, among others (F). Values in (E, F) are the negative log of the enrichment test P value after correction for multiple testing using the g:SCS algorithm in gProfilerR (Reimand et al. 2016). Inset in (F) shows average expression (log counts per million; cpm) across age for Vegfa in lowlanders (orange) versus highlanders (blue). ***P < 0.001. See supplementary table S4, Supplementary Material online, for full functional enrichment analysis results.

We performed WGCNA on skeletal muscle transcriptomes from lowland and highland deer mice sampled at postnatal days 14, 21, and 27. WGCNA identified seven modules ranging in size from 92 to 2,878 genes (table 2). A total of 3,529 out of 11,104 genes could not be assigned to any module (M0; supplementary table S2, Supplementary Material online). Although individual-level thermogenic capacity data do not exist for the samples sequenced, we did detect significant correlations between module eigengene and thermogenic capacity at the level of population and age for two modules, M2 (r = −0.94; P = 0.006) and M6 (r = 0.98; P < 0.001; fig. 4A and table 2). Despite the fact that we detected an overall effect of postnatal age, but not population, on both of these modules (table 2), expression followed a pattern that closely resembled population-level variation in thermogenic capacity. For module M6 in particular, as with thermogenic capacity (fig. 1B), expression was lower in highland compared with lowland deer mice at p14 (F_1,4 = 27.7; P = 0.006), but no different from lowlanders at p21 (F_1,5 = 0.6; P < 0.05) or p27 (F_1,6 = 0.6; P > 0.05; fig. 4B). The opposite was true of M2, whereby module expression at p14 was significantly higher in high-altitude mice (F_1,4 = 27.7; P = 0.02).

Table 2.

Association Test and ANOVA Results on Skeletal Muscle Modules.

Module	Module Size	Population		Age		Pop × Age		VO2 Correlation
		F 1,22	P	F 1,22	P	F 1,22	P	r	P
M1	92	3.4	0.6	15.0	0.006	2.4	1	0.18	0.99
M2	2,654	1.0	1	150.5	<0.001	0.3	1	−0.94	0.02
M3	1,052	11.9	0.02	1.8	1	2.3	1	−0.79	0.15
M4	288	97.8	<0.001	8.7	0.06	1.7	1	0.0002	0.99
M5	168	4.6	0.4	3.5	0.61	4.4	0.4	0.19	0.99
M6	2,878	3.2	0.7	64.4	<0.001	0.1	1	0.98	0.004
M7	443	91.1	<0.001	7.2	0.11	1.0	1	0.07	0.99

Note.—Effect of population (highland vs. lowland), age (postnatal day p14–p27), or their interaction on rank-transformed module expression (module eigengene values) is presented. Final two columns present results of Pearson correlation between module eigengene and thermogenic capacity. Note that correlations were performed on population-age means since VO₂ measurements did not exist for the individuals sequenced. P values from association tests and ANOVAs were corrected for multiple testing using the false-discovery rate method. Significant P-values are italicized.

Fig. 4.

Skeletal muscle module M6 is significantly associated with thermogenic capacity in lowland (orange, triangles) and highland (blue, circles) deer mice aged p14–p27. (A) Module expression is positively correlated with thermogenic capacity at the population/age level (individual-level data not available). (B) Highland deer mice exhibit significantly lower module eigengene values at postnatal day p14 (**P < 0.01), and no differences in expression at p21 or p27. (C) Functional enrichment analysis reveals enrichment of KEGG pathways involved in aerobic metabolism and thermogenesis among others (see also supplementary table S4, Supplementary Material online). Values are the negative log of the enrichment test P value after correction for multiple testing using the g:SCS algorithm in gProfilerR (Reimand et al. 2016). Expression values (log counts per million; cpm) for the transcription factors Esrra (D) and Ppara (E) at p14 are lower for highlanders compared with lowlanders; both transcription factors are involved in the regulation of mitochondrial biogenesis and metabolic pathways.

Functional enrichment analysis of VO₂-associated skeletal muscle modules revealed enrichment of a wide variety of biological functions (supplementary table S4, Supplementary Material online). Module M2 for example was enriched for pathways that function in the development of muscle blood supply, including the GO Biological Process terms “angiogenesis, blood vessel morphogenesis, and vasculogenesis.” This result may reflect an initiation of blood vessel development that eventually leads to a greater production of capillaries in highlanders by p21 (Robertson and McClelland 2019). Indeed, along with a greater proportion of oxidative fiber types, increased capillary density is thought to be an adaptive advantage at high altitude that contributes to improved aerobic performance capabilities in adult mice (Lui et al. 2015; Scott et al. 2015; Nikel et al. 2018).

By contrast to M2, module M6 was enriched for functions related to energy metabolism, generation of ATP, and thermogenesis, including the KEGG terms “oxidative phosphorylation, thermogenesis, citrate cycle, and glycolysis/gluconeogenesis” (fig. 4C). Enrichment of the GO Cellular Component terms “NADH dehydrogenase complex and mitochondrial respiratory chain complex I” indicates that energy metabolism functions are related to changes in the production of NADH: ubiquinone oxidoreductase/mitochondrial complex I, the first protein complex in the electron transport chain. We also detected significant enrichment of transcription factor binding sites for estrogen-related receptor alpha, ERR1 (supplementary table S4, Supplementary Material online; Esrra in the Peromyscus genome), which is a major regulator of mitochondrial biogenesis, gluconeogenesis, and oxidative phosphorylation. Indeed, the Esrra gene itself was assigned to module M6 and its expression at p14 is strongly downregulated in highlanders relative to lowlanders (fig. 4D). Detection of Ppara (peroxisome proliferator activated receptor alpha) in module M6 (fig. 4E) suggests that fatty acid oxidation may also be depressed in highlanders, since this gene is a major regulator of fat metabolism in muscle (Issemann and Green 1990).

Evidence of Selection on Candidate Transcriptional Modules

Among transcriptional modules identified by WGCNA, three (B3, B36, and M6) were significantly and positively associated with variation in VO₂. These modules represent candidate pathways underlying the delay in thermogenesis in high-altitude mice. We took a two-pronged approach to determine whether developmental delays in candidate module expression are adaptive at high altitude. First, we calculated P_ST (an estimate of phenotypic differentiation following Brommer [2011]) on module eigengene values of candidate modules to test whether divergence between lowland and highland populations exceeds the expectation set by neutral genetic differentiation (F_ST; estimated with Weir’s Theta; Weir and Cockerham 1984). P_ST was calculated across a range of c/h² values from 0 (no heritability) to 2, which is where the relationship between c/h² and P_ST reaches asymptote (Brommer 2011; fig. 5); the ratio c/h² represents the degree to which phenotypic differences are due to additive genetic effects and cannot be readily estimated using our data. For candidate BAT modules B3 and B36, P_ST was significantly greater (α = 0.05) than F_ST above extremely conservative estimates of c/h² that approached zero (B3: c/h² < 0.11; B36: c/h² < 0.08; fig. 5A and B). Using 99% confidence intervals raised the value of c/h² for which P_ST significantly exceeds F_ST, though this value remained at extremely conservative levels (B3: c/h² < 0.14; B36: c/h² < 0.15). We note that in all cases the values of c/h² for which P_ST is greater than F_ST are far lower than Brommer’s (2011) null expectation of c/h² = 1, suggesting that phenotypic differences are very unlikely to be caused by drift alone. This interpretation is further supported by the critical c/h², which for B3 and B36 are extremely low (<0.13; supplementary table S5, Supplementary Material online). By contrast to B3 and B36, confidence intervals for P_ST calculated on B0 (not associated with thermogenesis) overlap entirely with F_ST at all values of c/h² (fig. 5C), indicating that differentiation in expression of this module is nonsignificant and does not exceed neutral expectations. P_ST calculated on VO₂ across p0–p10 did not exceed the upper 95% limit of F_ST, except when we restricted it to p10 (supplementary fig. 2, Supplementary Material online), the time point under which cold-induced VO₂ is significantly lower in highlanders compared with lowlanders (fig. 1A).

Fig. 5.

P_ST on module eigengene values for BAT modules B3 (A), B36 (B), and B0 (C), which constitutes a suite of uncorrelated genes. P_ST values were calculated across a range of unknown c/h² values from 0 to 2. Gray polygons represent 95% confidence limits on P_ST. The upper 95% confidence of F_ST are shown as a dashed line. Ninety-five percent confidence intervals for both values were generated by bootstrapping. Nonoverlap between the lower limit of P_ST and upper F_ST confidence limits indicates a significant difference at α = 0.05. For nearly every value of P_ST for modules B3 and B36, but not B0, P_ST is significantly greater than F_ST.

We found that P_ST on the skeletal muscle module M6 did not exceed neutral expectations: although observed mean P_ST does exceed the upper 95% limit of F_ST, 95% confidence limits for both values overlap entirely with zero (supplementary fig. S1, Supplementary Material online). We note that P_ST does not tend to overlap with the upper 95% limit of F_ST when we restrict it to p14 (the time point of significant differentiation in module expression and thermogenic capacity; figs. 1B and 4B; see supplementary fig. S2, Supplementary Material online). This pattern of phenotypic differentiation is mirrored by P_ST on thermogenic capacity at p14, which greatly exceeds the upper 95% limit of F_ST at all values of c/h² (supplementary fig. S2, Supplementary Material online).

Finally, we tested whether these candidate VO₂-associated transcriptional modules were enriched for genes exhibiting signatures of positive natural selection, as measured by the population branch statistic (PBS; Yi et al. 2010). Roughly 25% of the genes in modules B3, B36, and M6 contained single-nucleotide polymorphisms (SNPs) with PBS values that exceeded neutral expectations (table 3). However, only B36 exhibited a significant enrichment of outlier SNPs relative to the background rate (Fisher’s exact test, P = 0.02; table 3). This analysis demonstrates that, although all phenotype-associated modules exhibit some proportion of genes under selection, only B36 contains more than was expected by chance.

Table 3.

Number and Proportion of Genes in Candidate Modules Containing Outlier SNPs (PBS > 99.9% quantile of the simulated distribution).

Module	No. Outlier Genes	Total Genes	% Outlier Genes	P Value
B3	184	773	23.8	0.7
B36	196	708	27.7	0.02
M6	631	2,878	21.9	1

Note.—P values obtained from Fisher’s exact tests. Significant P-values are italicized.

Discussion

We explored the mechanisms that underlie a developmental delay in the onset of thermogenesis in juvenile high-altitude deer mice (fig. 1). We show that the developmental trajectory of the thermoregulatory system has evolved at high altitude via selection on the regulatory systems that control the development of thermoeffector tissues, BAT and skeletal muscle. Combined with the results of Robertson et al. (2019) and Robertson and McClelland (2019), our work suggests that the cold hypoxic conditions of high altitude force an alternative resource allocation strategy, whereby limited energy is put into developmental processes such as growth, over the development of the thermoregulatory machinery needed in heat production.

A Developmental Delay in NST Is Associated with a Delay in Expression of Fuel Supply, Oxygenation, and Axon Guidance Pathways in BAT

During early postnatal development mice rely exclusively on BAT for thermogenesis, and low-altitude pups activate BAT to maintain body temperature in the face of cold by p10. Our analysis suggests that an ontogenetic delay in the development of NST in high-altitude deer mice (fig. 1A) is attributable to a delay in the nervous system innervation and O₂/fuel supply lines to a maturing BAT. WGCNA of BAT revealed two transcriptional modules (B3, B36) that are expressed at higher levels in lowlanders across age (fig. 3). Both modules were correlated with cold-induced VO₂, suggesting that changes in the expression of genes in these modules influence the development of thermogenesis. Module B3 was enriched for pathways involved in metabolism, particularly fatty acid degradation (fig. 3 and supplementary table S4, Supplementary Material online). This is compelling since fat is the primary fuel source that powers uncoupling of cellular respiration from ATP production in nonshivering heat production (Cannon and Nedergaard 2004). Indeed, the β3 signaling cascade, initiated by cold, stimulates the lipolysis of fats used in β-oxidation and the eventual uncoupling of ATP synthesis from electron transport via activation of Uncoupling Protein 1 (Ucp1). Although not assigned to B3, we found that expression of Ucp1 itself was lower in highlanders compared with lowlanders (analysis of variance [ANOVA], main effect of population; F_1,30 = 8.2; P = 0.007).

Module B36, by contrast, was enriched for two pathways that are important in O₂/fuel supply and cold-activation of BAT. For example, enrichment of the VEGF signaling pathway, and downregulation of Vegfa in highlanders (fig. 3B), suggests a reduction in the vascularization of BAT early in the postnatal period in high-altitude natives. This is notable since the VEGF cascade stimulates angiogenesis, increasing O₂ and metabolic fuel delivery. Moreover, Vegfa itself contains two SNPs above the 99.9% PBS threshold (Schweizer, Velotta et al. 2019), suggesting that changes in the VEGF pathway are one of many targets of selection at high altitude. We also detected enrichment of the KEGG pathway “axon guidance,” which represents a key stage in development in which neurons send out axons to reach their targets. Indeed, BAT is activated by the sympathetic nervous system in response to cold, which leads to the release of norepinephrine and the signaling cascade that produces heat (Cannon and Nedergaard 2004).

A reduction in the expression of genes that participate in axon guidance suggests a developmental delay in the sympathetic innervation of BAT, which would inhibit recruitment and activation of this tissue during cold exposure (Robertson et al. 2019). Robertson et al. (2019) found lower levels of the enzyme tyrosine hydroxylase in the BAT of high-altitude pups at p10, lending support to the hypothesis that neurotransmitter synthesis and neural activation of BAT tissue is delayed in highlanders; tyrosine hydroxylase is the rate-limiting enzyme in norepinephrine production and is present in sympathetic neurons innervating BAT. Related to the reduced sympathetic activation of BAT, we found that expression of the gene calsyntenin 3 (Clstn3) is also downregulated in highlanders (ANOVA, main effect of population; F_1,30 =15.1; P < 0.001). Clstn3 is a known promoter of synapse formation; a recent study in house mice found that a mammalian-specific form (Clstn3β) enhances functional sympathetic innervation of BAT (Zeng et al. 2019).

Despite downregulation of genes involved in O₂/fuel supply and innervation of BAT, Robertson et al. (2019) found that BAT mass does not differ between highlanders and lowlanders, and that citrate synthase (a biomarker of mitochondrial abundance) and UCP protein expression are likewise equivalent. Our results therefore suggest that, although BAT growth and metabolic potential may be equal in highlanders and lowlanders, a delay in the development of the innervation and O₂/fuel supply to BAT may lead to an inability of highlanders to respond to cold temperatures and thus mount an appropriate thermoregulatory response.

A Developmental Delay in Thermogenic Capacity Is Associated with Downregulation of Energy Metabolism Pathways in Muscle

Transcriptome analysis of skeletal muscle revealed a large transcriptional module (M6) that closely tracked variation in thermogenic capacity (fig. 4). This result suggests that the genes expressed in module M6 underlies the developmental delay in thermogenesis beyond p10 (fig. 1B). The M6 module (table 2) was enriched for a host of functions related to skeletal muscle metabolism that reflect changes in its maturation with respect to shivering. Our analysis suggests that high-altitude mice show a developmental delay in the expression of nearly all pathways related to ATP generation by mitochondria, including glycolysis, citrate cycle, and oxidative phosphorylation (fig. 4C). Many of the genes that contribute to functional enrichment of metabolic pathways were related to the production of mitochondrial respiratory complex I, including 31 of 35 Nduf genes, suggesting that downregulation of metabolic pathways may be related to a delay in mitochondrial biogenesis. Indeed, the GO terms “mitochondria and mitochondrial part” were also enriched (supplementary table S4, Supplementary Material online).

At least two transcription factors that are key regulators of cellular metabolism were found in module M6 and were expressed at lower levels in highlanders (fig. 4D and E). Estrogen-related receptor alpha, Esrra, for example, directs the expression of genes involved in mitochondrial biogenesis (Wu et al. 1999) and mitochondrial energy-producing pathways in skeletal muscle (Huss et al. 2004). Genes in module M6 were also enriched for Esrra binding sites (supplementary table S4, Supplementary Material online). Lowered expression of peroxisome proliferator activated receptor alpha (Ppara) suggests that fatty acid oxidation may also be depressed in highlanders, since this gene is a major regulator of fat metabolism capacity; functional terms related to fat oxidation however were not enriched in this module (supplementary table S4, Supplementary Material online). Indeed, Robertson and McClelland (2019) found that the β-oxidation enzyme β-hydroxyacyl-CoA dehydrogenase did not track changes in muscle aerobic capacity of highlanders over these ages. These data provide evidence that the developmental delay in thermogenic capacity is linked to a delay in the expression of pathways that facilitate mitochondrial development and muscle aerobic capacity and allow for shivering thermogenesis.

Evidence of Natural Selection

We found that population divergence in expression of VO₂-associated transcriptional modules exceeded neutral expectations set by genetic differentiation (measured as F_ST), suggesting that population differences may be adaptive. We expected that, if population divergence in module expression was adaptive, P_ST on module eigengene would exceed the upper bounds of the confidence intervals around our estimate of F_ST. At extremely low and conservative values of c/h² (i.e., c/h² < 1, the null expectation according to Brommer [2011]), the lower confidence limits of P_ST for candidate BAT modules B3 and B36 exceeded the upper confidence limits of F_ST (fig. 5A and B), indicating significant differences. These results hold true when 95% and 99% confidence intervals are used, suggesting that population-level expression differences in candidate BAT modules are highly unlikely to be caused by drift alone. This finding is consistent with P_ST on cold-induced VO₂ values at p10 (supplementary fig. S2, Supplementary Material online), the time point of greatest population differentiation (fig. 1A). Although P_ST did not exceed F_ST for the skeletal muscle module M6 across all time points (p14–p27), we did find it to exceed neutral expectations when restricted to p14 (supplementary fig. S2, Supplementary Material online). This is consistent with the finding that P_ST on thermogenic capacity also exceeds the upper 95% of F_ST at p14 only (supplementary fig. S2, Supplementary Material online). Together these data suggest that the delay in the onset of NST and shivering thermogenesis, and their underlying regulatory control pathways, may be the target of selection at high altitude.

Approximately 25% of genes in candidate modules bear the signature of natural selection, that is, they contain at least one SNP above the PBS significance threshold (table 3). Only module B36, however, was significantly enriched for PBS outliers above the transcriptome-wide background level. Such sequence divergence within candidate module genes may be the result of selection at the cis-regulatory regions that modulate individual gene expression. However, that we do not detect a higher proportion of candidate module genes bearing the signature of natural selection is not unexpected, as selection on upstream trans-regulatory elements may influence the expression of many genes in a single network. We note that it is not possible to discern with certainty whether the genetic variants we identified are the direct targets of selection, as these loci may be linked to casual variants. Pinpointing the genetic variants that influence candidate pathway expression and the delay in the development of thermogenesis will be the focus of future work. Nevertheless, because PBS is polarized by the use of a second lowland population (see Materials and Methods), we can discern that the outlier loci we do detect (or closely linked loci) are under positive selection in the high-altitude population. Our results support the interpretation that differentiation in expression of candidate modules, at least within BAT, is driven, in part, by natural selection.

Summary and Conclusions

We found that a developmental delay in the onset of independent thermoregulatory ability in deer mice native to high altitude is rooted in a delay in the expression of gene regulatory networks that contribute to sympathetic innervation of BAT that permits the response to cold, and in pathways that contribute to fuel use and ATP production of both thermoeffector organs. Phenotypic divergence analysis suggests that both thermogenic delays and associated shifts in gene expression in BAT are driven by natural selection. An overabundance of genes exhibiting signatures of natural selection in BAT module B36 suggests that developmental changes to genes in this module may be particularly important in high-altitude adaptation. Our results demonstrate that changes to the developmental trajectory of thermoregulation at high altitude are the result of regulatory delays in the development of thermoeffector organs, which limit the ability of high-altitude mice to mount a thermoregulatory response.

Our results support the hypothesis that the developmental delay in thermogenesis at high altitude is attributable to an adaptive energetic tradeoff (Robertson et al. 2019; Robertson and McClelland 2019). The observed developmental delay in expression of transcriptional pathways associated with neural activation and energy production suggests that the BAT and skeletal muscles of high-altitude mice take longer to fully mature postnatally, relative to lowlanders. That growth rate is equivalent between high- and low-altitude pups under common conditions (Robertson et al. 2019) suggests that high-altitude mice may allocate energy toward growth instead of into the maturation of thermoeffector tissues that promote active thermogeneration. The combination of chronic cold and low O₂ availability at high altitude, coupled with larger litter sizes that should affect competition among siblings for limited resources (Robertson et al. 2019), may contribute to the altered energetic conditions that drive this tradeoff. Because high-altitude adult deer mice show the opposite pattern compared with developing pups (i.e., consistently higher thermogenic capacities, Cheviron et al. 2013), we suggest that selection pressures at high altitude can have very different effects on the same physiological systems, depending on age. Future work on the selective drivers and energetic benefits of this putatively adaptive energy allocation strategy will shed light on adaptive modification of developmental processes to meet the challenge of selection pressures that vary in strength and magnitude across ontogeny.

Materials and Methods

Animals and Experimental Procedures

Deer mice used in this study were F₂ lab-reared descendants of two wild populations of Peromyscus maniculatus (Robertson et al. 2019; Robertson and McClelland 2019) raised under common environmental conditions (24 °C, 80 m a.s.l.) at McMaster University of Ontario, Canada. Individuals from the high-altitude population were captured from the summit of Mount Evans in Clear Creek County, CO, USA (4,350 m a.s.l.) and the low-altitude population from Nine-mile prairie, NE, USA (320 m a.s.l.). F₂ pups were derived from one of the first three litters produced by their F₁ parental breeding pair (family). A total of 7 lowland and 11 highland families were used. At each developmental age, two individuals from each family (one male and one female, when possible) were sampled. Sampling time points were randomly distributed across the three litters.

Altricial rodents, such as deer mice, usually develop NST prior to shivering due to the maturation of BAT which precedes the maturation of skeletal muscle. P. maniculatus are only capable of NST after the first 10 days postpartum (Robertson et al. 2019). In contrast, shivering develops between 2 and 4 weeks after birth (Robertson and McClelland 2019). We sampled BAT and skeletal muscle (specifically, the gastrocnemius) over these two periods. Briefly, pups were removed from the nest, then euthanized with an overdose of isoflurane followed by cervical dislocation. In pups ages 0, 2, 4, 6, 8, and 10 days postpartum the intrascapular depot of BAT was blunt dissected and cleaned of white adipose tissue. In pups aged 14, 21, and 27 days postpartum the gastrocnemius muscle of the lower hindlimb was blunt dissected. Tissues were flash frozen and stored at −80 °C.

We reanalyzed cold-induced O₂ consumption rate (VO₂) data from Robertson et al. (2019) and Robertson and McClelland (2019) for low- and high-altitude native P. maniculatus only, omitting the low-altitude congeneric, Peromyscus leucopus. Briefly, for pups aged 0–10 days postpartum cold-induced VO₂ was assessed in response to a mild cold stress (10 min at 24 °C). Pups were compared with a control litter mate who was maintained at 30 °C for the same duration of trial in order to control for handling stress. For pups aged 14–27 days, maximum cold-induced VO₂ (thermogenic capacity) was measured using previously established methods for adult deer mice (Cheviron et al. 2012). VO₂max was induced by exposing pups to −5 °C in heliox air (21% O₂ with He). All VO₂ measurements were divided by mass to obtain mass-specific metabolic rates.

Transcriptome Analyses

We conducted high-throughput sequencing of RNA of BAT and skeletal muscle in order to explore the mechanisms that underlie developmental delays in thermogenesis. Sample size varied from n = 1–7 for each population, age, and tissue (see supplementary table S1, Supplementary Material online, in online supplementary material for full list). We assayed gene expression using TagSeq, a 3′ tag-based sequencing method following Lohman et al. (2016). First, we extracted RNA from <25 mg of tissue using TRI Reagent (Sigma-Aldrich), then assessed RNA quality using TapeStation (Agilent Technologies; RNA Integrity Number > 7). The Genome Sequencing and Analysis Facility at the University of Texas at Austin prepared TagSeq libraries, which were sequenced using Illumina HiSeq 2500. We filtered raw reads for length, quality, and polymerase chain reaction duplicates following Lohman et al. (2016) using scripts modified from those available online (https://github.com/z0on/tag-based_RNAseq; last accessed April 16, 2020). Using the FASTX-toolkit (http://hannonlab.cshl.edu/fastx_toolkit/; last accessed April 16, 2020), we cleaned and trimmed raw reads, which resulted in 3.2 million reads per individual. Filtered reads were then mapped to the P. maniculatus genome (NCBI GCA_000500345.1 Pman_1.0) using bwa mem (Li and Durbin 2010). Across all samples, ∼64–79% of raw reads mapped to the P. maniculatus genome. We detected 3% lowered mapping success, on average, in lowland individuals in the BAT transcriptome (linear mixed effects model; F_1,8.2 = 5.6; P = 0.04), but no population difference in the skeletal muscle transcriptome (P > 0.05). We used featureCounts (Liao et al. 2014) to generate a table of transcript abundances. Since genes with low read counts are subject to measurement error (Robinson and Smyth 2007), we excluded those with less than an average of 10 reads per individual. We retained a total of 11,192 and 11,104 genes after filtering for BAT and skeletal muscle transcriptomes, respectively. Five outlier samples (one BAT, four skeletal muscle) were removed following visual inspection of multidimensional scaling plots using plotMDS in edgeR following Chen et al. (2019).

We assessed overall patterns of gene expression among BAT and skeletal muscle transcripts using PCA, whereas WGCNA (v. 1.41-1; Langfelder and Horvath 2008) was used to identify potential regulatory mechanisms that underlie variation in thermogenesis across development. PCA and WGCNA analyses were conducted separately on BAT and skeletal muscle samples in R v. 2.5.2 (R Core Team). WGCNA identified clusters of genes with highly correlated expression profiles (modules). This approach was successfully implemented in recent studies that aimed to relate gene expression with higher-level phenotypes (Plachetzki et al. 2014; Velotta et al. 2016, 2018). Prior to performing WGCNA, we normalized raw read counts by library size and log-transformed them using the functions calcNormFactors and cpm, respectively, from the R package edgeR (Robinson et al. 2010). Module detection was performed using the blockwiseModules function in WGCNA with networkType set to “signed” (Langfelder and Horvath 2008). Briefly, Pearson correlations of transcript abundance data were calculated between pairs of genes, after which an adjacency matrix was computed by raising the correlation matrix to a soft thresholding power of β = 6. Soft thresholding power β is chosen to achieve an approximately scale free topology, an approach that favors strong correlations (Zhang and Horvath 2005). We chose a β = 6 since it represents the value for which improvement of scale free topology model fit begins to decrease with increasing thresholding power. A topological overlap measure was computed from the resulting adjacency matrix for each gene pair. Topologically based dissimilarity was then calculated and used as input for average linkage hierarchical clustering in the creation of cluster dendrograms for both tissues. Modules were identified as branches of the resulting cluster tree using the dynamic tree-cutting method (Langfelder and Horvath 2008). We assigned modules a unique identification according to tissue (B0–B37 for BAT; M0–M7 for skeletal muscle). Genes that could not be clustered into a module were given the designation B0 and M0 for BAT and skeletal muscle, respectively.

Once modules were defined, we used a multistep process to associate expression among individuals with variation in VO₂. First, we summarized module expression using PCA of gene expression profiles (blockwiseModules in WGCNA); because genes within modules are highly correlated by definition, the first principal component axis (referred to as the module eigengene) was used to represent module expression (Langfelder and Horvath 2008). We used module eigengene values to test for associations between module expression and VO₂ for each module (Pearson correlation; cor function in WGCNA) in the BAT network. P values for the correlation were determined by a Student’s asymptotic test (corPvalueStudent in WGCNA). For skeletal muscle, association tests were conducted on population-age means since VO₂ measurements do not exist for the individuals sequenced. Among phenotype-associated modules, we conducted analysis-of-variance (ANOVA) on rank-transformed module eigengene values in order to test for the effects of population, age, and their interaction on module expression. P values from association tests and ANOVAs were corrected for multiple testing using the false-discovery rate method (Benjamini and Hochberg 1995).

We performed functional enrichment analysis on all modules exhibiting either a statistical association with VO₂, or a significant effect in ANOVA, or both, using the R package gProfilerR (Reimand et al. 2016). P. maniculatus gene names were converted to Mus musculus gene names (Online supplementary material) using a custom script. We corrected for multiple testing using gProfilerR’s native g: SCS algorithm. We used the list of genes from filtered BAT and skeletal muscles transcriptomes as a custom background list in all enrichment analyses. We searched for enrichment among terms from GO (Ashburner et al. 2000; Gene Ontology Consortium 2019), KEGG (Kanehisa and Goto 2000), Reactome (Fabregat et al. 2018), and WikiPathways (Slenter et al. 2018) databases, as well as transcription factor binding sites from the TRANSFAC database (Matys et al. 2006).

Tests of Phenotypic and Genetic Divergence

We used two complementary approaches to assess phenotypic and genetic divergence to test for signatures of natural selection in VO₂-associated modules. To assess phenotypic divergence, we calculated P_ST on module eigengene values following Brommer (2011) using equation (1):

$$P_{\mathrm{ST}}=\frac{\frac{c}{h^2}{\sigma }_{\mathrm{B}}^2}{\frac{c}{h^2}{\sigma }_{\mathrm{B}}^2+2{\sigma }_{\mathrm{W}}^2},$$ (1)

where ${\sigma }_{\mathrm{B}}^2$ denotes between-population phenotypic variance, ${\sigma }_{\mathrm{W}}^2$ denotes within-population phenotypic variance, h² the narrow-sense heritability (the proportion of phenotypic variance that is due to additive genetic effects), and c a scalar representing the total phenotypic variance due to additive genetic effects across populations. Within- and between-population variances were estimated from ANOVA, where module eigengene values were response variables, and postnatal day, population, and family nested within population were included as main effects. Within-population variance (${\sigma }_{\mathrm{W}}^2)$ was calculated as the residual mean squares (MS) from the ANOVA model. Between-population variance was estimated after Antoniazza et al. (2010) following equation (2):

$${\sigma }_{\mathrm{B}}^2=\frac{{\mathrm{MS}}_{\mathrm{B}}-{\mathrm{MS}}_{\mathrm{W}}}{n_0},$$ (2)

where MS_B and MS_W are the population effect and the residual mean square variances from the ANOVA model, respectively, and n₀ a weighted average of sample size based on Sokal and Rohlf (1995) following equation (3):

$$n_0=\frac{\left(n_1+n_2\right)-({n1}^2+{n_2}^2)}{n_1+n_2},$$ (3)

where n₁ and n₂ are the sample sizes of population one and two, respectively. The ratio c/h² was unknown and could not be estimated using our experimental design. As such, we calculated P_ST across a range of values of c/h² from 0 (no heritability) to 2, which is where the relationship between c/h² and P_ST begins to reach asymptote (Brommer 2011; fig. 5). We calculated 95% and 99% confidence intervals for each value of P_ST calculated across values of c/h² ranging from 0 to 2 at 0.05 level increments (total of 40 P_ST estimates). Confidence intervals for each P_ST value were estimated with bootstrapping (with 1,000 replicates) using the boot function in R.

P _ST estimates were compared with neutral divergence expectations set by between-population differentiation. We calculated pairwise genetic differentiation between the lowland and highland populations using F_ST (estimated with Weir’s Theta; Weir and Cockerham 1984) and previously sequenced exome data (Schweizer, Velotta et al. 2019). Briefly, F_ST was calculated for 5,183,435 high-quality biallelic SNPs using the “–weir-fst-pop” flag within vcftools v0.1.16 (Danecek et al. 2011). We pruned SNPs for linkage-disequilibrium using the “indep-pairwise 50 5 0.5” flag in PLINK1.9 (Chang et al. 2015) and restricted the data to nongenic, putatively neutral, regions of the exome (Schweizer, Velotta et al. 2019), leading to a reduced set of 172,408 SNPs. We then calculated 95% and 99% confidence intervals around the estimate of mean F_ST (0.022) with bootstrapping by randomly sampling a subset of 10,000 SNPs, with replacement, 10,000 times (boot function in R). Significant differences between P_ST and F_ST were determined by nonoverlap between the upper confidence limits of F_ST and lower limits of P_ST for α = 0.05 for 95% confidence limits, and α = 0.01 for 99% confidence limits. Finally, we calculated the critical c/h² value following Brommer (2011), which represents the ratio at which confidence intervals around P_ST equal the confidence interval around F_ST. These results are presented in supplementary table S5, Supplementary Material online, for ease of comparison to other published quantitative analyses.

We then compared P_ST of each candidate VO₂-associated module with that of “module 0,” which contains the suite of genes that could not be reliably clustered (B0 in BAT; M0 in skeletal muscle). The null expectation is that P_ST should not exceed the neutral expectation set by F_ST at any value of c/h². We expected that, if natural selection has shaped shifts in gene expression, then P_ST would exceed F_ST in VO₂-associated modules, but not in modules B0 or M0 (which serve as a measure of background levels of expression differentiation). We calculated P_ST on cold-induced VO₂ (p0–p10) and thermogenic capacity (p14–p27) as above in order to determine whether these higher-level phenotypic traits were also the targets of natural selection.

Finally, we took a population genomic approach to assess whether genes in VO₂-associated transcriptional modules exhibit signatures of positive natural selection at high altitude. To do this, we used PBS (Yi et al. 2010) data from Schweizer, Velotta et al. (2019). Here, PBS identifies loci that exhibit extreme allele frequency differences in highland (Mt Evans) relative to lowland (Lincoln and Merced, CA) populations. We previously calculated PBS for 5,183,435 exome-wide biallelic SNPs, then calculated the demographically corrected significance of PBS values using a simulated distribution of 500,000 SNPs (see Schweizer, Velotta et al. 2019 for simulation details). We first determined significant outlier SNPs with a PBS value located above the 99.9% quantile, then identified those SNPs located within genes belonging to VO₂-associated modules using Ensembl’s Variant Effect Predictor (McLaren et al. 2010) with the P. maniculatus reference genome annotation data set. We used Fisher’s exact tests to determine whether thermogenesis-associated modules were statistically enriched for loci that exceeded the 99.9% threshold.