Signatures of positive selection and local adaptation to urbanization in white-footed mice (Peromyscus leucopus)

Abstract

Urbanization significantly alters natural ecosystems and has accelerated globally. Urban wildlife populations are often highly fragmented by human infrastructure, and isolated populations may adapt in response to local urban pressures. However, relatively few studies have identified genomic signatures of adaptation in urban animals. We used a landscape genomics approach to examine signatures of selection in urban populations of white-footed mice (Peromyscus leucopus) in New York City. We analyzed 154,770 SNPs identified from transcriptome data from 48 P. leucopus individuals from three urban and three rural populations, and used outlier tests to identify evidence of urban adaptation. We accounted for demography by simulating a neutral SNP dataset under an inferred demographic history as a null model for outlier analysis. We also tested whether candidate genes were associated with environmental variables related to urbanization. In total, we detected 381 outlier loci and after stringent filtering, identified and annotated 19 candidate loci. Many of the candidate genes were involved in metabolic processes, and have well-established roles in metabolizing lipids and carbohydrates. Our results indicate that white-footed mice in NYC are adapting at the biomolecular level to local selective pressures in urban habitats. Annotation of outlier loci suggest selection is acting on metabolic pathways in urban populations, likely related to novel diets in cities that differ from diets in less disturbed areas.

INTRODUCTION

Urban habitats are one of the fastest growing and most rapidly changing environments around the world. While urbanization has been traditionally viewed as a driver of declining habitat quality in and around cities, there is growing interest in the idea that urban areas represent novel environments with unique selective pressures (Donihue & Lambert 2015). The recently developed but burgeoning field of urban evolutionary biology aims to determine how urbanization leads to evolutionary change through mutation, genetic drift, gene flow, and natural selection in urban populations.

The ecological changes that occur within cities are likely to have many evolutionary implications. Human infrastructure causes habitat loss and fragmentation and changes resource availability, novel species interactions occur because human movements and commerce introduce a diverse array of nonnative species, and human activity increases exposure to chemical, light, and noise pollution (McKinney 2002; Chace & Walsh 2004; Shochat et al. 2006; Sih et al. 2011). These changes lead to unique pressures in novel urban habitats that may rapidly drive evolutionary change over short timescales. Increased genetic drift in relatively isolated urban populations, genetic differentiation between populations with restricted gene flow from urban infrastructure, or allele frequency shifts due to local urban adaptation, are all likely outcomes of evolution in cities (Munshi-South 2012; Merilä & Hendry 2014; Donihue & Lambert 2015).

Urban populations are potentially excellent systems for examining how species respond to anthropogenic environmental change, what genes and traits are involved, and how quickly populations locally adapt to changing environments. Local adaptation is a common phenomenon in nature (Stinchcombe & Hoekstra 2008; Bonin 2008; Linnen et al. 2009; Hohenlohe et al. 2010a; Turner et al. 2010; Ellison et al. 2011; De Wit & Palumbi 2013), and often results from the operation of selection on standing genetic variation as opposed to novel mutations over relatively short time scales (Barrett & Schluter 2008; Stapley et al. 2010). Additionally, the quantitative traits involved in local adaptation may involve many genes of small effect working to produce the desired phenotype (Orr 2005; Rockman 2012), and these ecologically relevant but non-conspicuous phenotypes are predicted to be those most involved in urban adaptation (Sih et al. 2011). However, traits with relatively simple genetic architecture may also be under selection in urban environments (Thompson et al. 2016). Investigating the genetic basis of local adaptation has provided insight into a variety of evolutionary processes including speciation, maintenance of genetic diversity, range expansion, and species responses to changing environments (Savolainen et al. 2013; Tiffin & Ross-Ibarra 2014), and holds great promise for understanding adaptive evolution in response to urbanization.

Landscape genomics has recently produced a number of approaches for studying local adaptation. This field is defined by the spatially explicit study of genomic variation (Sork et al. 2013) that seeks to identify environmental variables influencing adaptive genomic variation (Rellstab et al. 2015). Landscape genomics, and more specifically genotype-by-environment analyses (GEA), can successfully identify associations between urban environmental variables and allele frequencies that indicate adaptation to local urban conditions. These approaches can also help to untangle the interactions between neutral demographic processes and selection (Rellstab et al. 2017). Urban populations are influenced by both genetic drift through founder effects and barriers to gene flow, and selection acting on genetic variation linked to increased fitness in urban settings.

A small but growing number of studies have documented how populations may locally adapt to urban selective pressures through changes in allele frequencies and/or undergo directional shifts in phenotypic traits. Yeh (2004) reported that sexually-selected tail coloration in Juncos (Junco hyemalis) was rapidly evolving in urban populations compared to rural ones. European Blackbirds (Turdus merula) exhibit evidence of selection on genes underlying anxiety behavior in newly established populations across multiple cities (Partecke et al. 2006; Mueller et al. 2013). Cheptou et al. (2008) reported that a weed (Crepis sancta) in urban vegetation plots surrounded by paved surfaces showed heritable changes in seed morphology and dispersal. Reduced snow cover in urban areas leads to colder minimum ground temperatures and Thompson et al. (2016) found parallel adaptive evolution in urban white clover (Trifolium repens) populations that had increased freezing tolerance. Several studies have also found likely adaptive genetic and morphological changes in urban mammal populations. Suggestive of urban adaptation, a specific mitochondrial genotype rose to fixation in white-footed mice (Peromyscus leucopus) populations in Chicago along with morphological changes to skull shape after urbanization (Pergams & Lacy 2008). In urban areas of Italy, Kuhl’s pipistrelle (Pipistrellus kuhlii) bat populations had significantly larger bodies and longer skulls than natural populations, suggesting urban adaption to a novel diet introduced when artificial illumination attracted an increased number of large hard-bodied moths (Tomassini et al. 2014).

Few studies in urban evolutionary biology have been able to measure phenotypic changes, definitively link them to genetic changes, and establish fitness benefits to demonstrate evolutionary adaptation. One exception are urban killifish (Fundulus heteroclitus), where selective pressure from polychlorinated biphenyls (PCBs) has led to the evolution of PCB tolerance in urban populations (Whitehead et al. 2010; Reid et al. 2016). Adaptation to PCB pollution was also reported in tomcod (Microgadus tomcod) in the Hudson River through a deletion that similarly increases tolerance to PCBs (Wirgin et al. 2011). Urban adaptation has also been confirmed in the well-known peppered moth (Biston betularia) system. Recent evidence suggests that the industrial melanism trait in this species is linked to an insertion of a transposable element in the cortex gene in the early 1800s that spread throughout the population in response to industrial airborne pollution (Hof et al. 2016). The study of additional systems will likely identify a complex array of adaptive evolutionary responses in cities (Whitehead et al. 2017).

Here we examined signatures of selection in isolated urban populations of white-footed mice, Peromyscus leucopus, in New York City (NYC) using a landscape genomics approach. Peromyscus spp. (Rodentia, Cricetidae) are a group of abundant small mammals found across much of North and Central America. They live in a diverse array of habitats that exposes them to a variety of selective pressures, and thus multiple Peromyscus spp. have become model systems for studies examining ecology, evolution, and physiology in natural populations (Munshi-South & Richardson 2017). There is also evidence that Peromyscus spp. readily adapt to environmental change (Storz et al. 2007, 2009, 2010; Mullen & Hoekstra 2008; Linnen et al. 2009; Weber et al. 2013; Natarajan et al. 2013; Munshi-South & Richardson 2017), making them good subjects for the study of local adaptation. White-footed mice are one of the few native mammals that thrive in extremely small, fragmented urban forests in North America (Pergams & Lacy 2008; Rogic et al. 2013; Munshi-South & Nagy 2014), and tend to be found at higher densities in urban vs. rural patches due to a thick understory providing abundant food resources and exclusion of major predators and competitors (Rytwinski & Fahrig 2007). Increased density may also be due to limited P. leucopus dispersal between urban sites. Munshi-South (2012) found barriers to dispersal between isolated NYC parks, with migrants only moving through significantly vegetated corridors throughout the city. There is also substantial genetic structure between NYC parks as measured by microsatellites (Munshi-South & Kharchenko 2010), genome-wide SNPs (Munshi-South et al. 2016) and demographic modeling (Harris et al. 2016). We have also previously identified signatures of selection in urban populations of NYC white-footed mice (Harris et al. 2013), though we used smaller datasets and more limited approaches than presented here.

In the current study, we examined SNPs generated from individual transcriptome sequencing for P. leucopus from three urban sites in NYC and three rural sites from the surrounding area. We generated a large SNP dataset and produced estimates of nucleotide diversity (π, Tajima 1983), Tajima’s D (Tajima 1989), and F_ST (Wright 1951) to generate per-site estimates and identify loci that deviate from neutral expectations. We then used a variety of genome scan methods and outlier tests to identify genes subject to selection in an urban setting. Our approach identified population differentiation, shifts in allele frequencies, and associations between alleles and environmental variables. However, neutral demographic processes such as population bottlenecks can produce signatures of genetic variation similar to those produced by selection (Oleksyk et al. 2010; Li et al. 2012). We accounted for this possibility by incorporating a simulated neutral SNP dataset from an inferred demographic history (Harris et al. 2016) directly into our null model for identifying outliers (Excoffier et al. 2009; Gutenkunst et al. 2009; Li et al. 2012; Vitti et al. 2013; Lotterhos & Whitlock 2015).

The three specific aims of this study were the following: 1. identify candidate genes exhibiting signatures of selection in NYC populations of white-footed mice using a variety of genome scan methods and outlier tests; 2. distinguish genetic outliers resulting from selection rather than demography by incorporating demographic histories of white-footed mice in NYC into null models of genome scans; and 3. identify genes that are statistically associated with environmental variables representative of urbanization using landscape genomic approaches.

MATERIALS AND METHODS

Sampling, library preparation, and transcriptome assembly

We trapped and collected white-footed mice from 2010 – 2012. For full details on sampling, transcriptome sequencing, assembly and SNP calling, see Harris et al. 2013, 2015. In brief, we randomly chose eight adult white-footed mice (equal numbers of males and females) from each of six sampling locations (N = 48 total) representative of urban and rural habitats and with minimal within-site genetic structure (Fig. 1) (Harris et al. 2013, 2015). Three sampling sites were within NYC parks: Central Park in Manhattan (CP), New York Botanical Gardens in the Bronx (NYBG), and Flushing Meadows—Willow Lake in Queens (FM). These sites represented urban habitats surrounded by high levels of impervious surface cover and high human population density, as previously quantified in Munshi-South et al. (2016). The remaining three sites occurred ~100 km outside of NYC in rural, undisturbed habitat representative of natural environments for P. leucopus. High Point State Park is in the Kittatinny Mountains in New Jersey (HIP), Clarence Fahnestock State Park is located in the Hudson Highlands in New York (CFP), and Brookhaven and Wildwood State Parks occur on the northeastern end of Long Island, New York (BHWWP). Total RNA was extracted separately from livers stored in RNA later for each of the 48 mice, treated with DNase, enriched through ribosomal RNA depletion, fragmented, reverse transcribed, amplified and tagged with a unique barcode, and sequenced in four lanes of one SOLiD 5500XL run (Harris et al. 2015). We called SNPs with the Genome Analysis Toolkit (GATK version 2.8) pipeline using a Bayesian genotype likelihood model (DePristo et al. 2011). In order to call a SNP, we required it to occur in at least five individuals, have a nucleotide quality (q-score) ≥ 30, exhibit no strand bias (FS ≥ 35), and to come from only uniquely mapped reads. We also required SNPs to have an overall depth ≥ 10X and ≤ 350X (to account for paralogous sequences), a minor allele frequency (MAF) ≥ 0.025, and removed SNPs where every individual was heterozygous.

Figure 1.

Summary statistics

SNP information was stored in a VCF (variant call format) file and summary statistics were calculated using vcftools 0.1.12b (Danecek et al. 2011). We calculated per-site nucleotide diversity (π), Tajima’s D, and F_ST. We also calculated the statistics for each contig (per-site statistic summed across all SNPs per contig divided by total sites) and calculated the average estimate for each population, including all pairwise population comparisons for F_ST.

Scans for positive selection based on population differentiation

We used the F_ST based analysis implemented in BayeScan v. 2.1 (Foll & Gaggiotti 2008) to compare all six population-specific allele frequencies with global averages and identify outlier SNPs. BayeScan identifies loci that exhibit divergence between groups that is stronger than would be expected under neutral genetic processes. Based on a set of neutral allele frequencies under a Dirichlet distribution, BayeScan uses a Bayesian model to estimate the probability that a given locus has been subject to selection. To generate more realistic allele frequency distributions, we used BayeScan for independent coalescent simulations of SNP datasets based on a neutral demographic history inferred by Harris et al. (2016) specifically for each P. leucopus population. Using the coalescent-based fastsimcoal2 software (Excoffier et al. 2013), we generated 100 sets of 100,000 SNPs each for every population in this study from a three population isolation-with-migration model using parameter estimates for divergence time, effective population size, migration rate, and population size change previously inferred in Harris et al. (2016). In short, the model represented a deep split between an ancestral population into Long Island, NY and the mainland (including Manhattan) 29,440 generations before present (GBP). A third population (representing the sampling sites in this study) later became isolated 746 GBP. Urban populations were also modeled to include a population size change event at the time of divergence. BayeScan was run independently on each of the 100 simulated datasets from fastsimcoal2 using default parameters to generate a null distribution of BayeScan statistics.

BayeScan was then run on the observed SNP dataset using default parameters. We performed several different analyses including a global analysis, one with two populations representing urban and rural groups, and finally analyses on all sampling site pairwise comparisons. We retained outlier SNPs with a q-value ≤ 0.1 (leading to a FDR of ≤ 0.1) and with a posterior odds probability from BayeScan higher than for any value calculated from the simulated dataset. BayeScan also calculates alpha (α), a locus specific F_st coefficient, where a positive value suggests diversifying selection and a negative value suggests balancing or purifying selection. There were no SNPs with negative α values.

For comparison to BayeScan results, we used a related method, BayPass (Gautier 2015), that identifies loci subject to selection based on allele frequency patterns that deviate from neutral expectations. We ran BayPass using default parameters under the auxillary covariate (AUX) model, and simulated pseudo-observed datasets (PODs) under the Inference Model in Baypass as suggested by Gautier (2015) to calibrate neutral distributions for XtX. BayPass uses the XtX statistic to identify adaptive divergence. SNPs with XtX estimates greater than the 95% threshold determined from PODs were identified as resulting from adaptive divergence.

Analysis for selective sweeps

We also identified outlier regions when the observed SFS showed an excess of low frequency and high frequency minor alleles, a signal indicative of a recent selective sweep. The composite likelihood ratio (CLR) statistic is used to identify regions where the observed SFS matches the expected SFS generated from a selective sweep (Kim & Stephan 2002; Nielsen et al. 2005; Pavlidis et al. 2010). We calculated the CLR along sliding windows across the transcriptome using the software program SweeD (Pavlidis et al. 2013). SweeD is an extension of Sweepfinder (Nielsen et al. 2005) that is optimized for large next generation sequencing (NGS) datasets. We lacked a genome to provide high-quality linkage information so SweeD was run separately for each population and on individual contigs. We used default parameters except for using a sliding window size of 200 bp and use of a folded SFS, as we lacked an outgroup to infer ancestral alleles. The window within each contig with the highest CLR score is considered the likely location of a selective sweep. Similar to the method used for BayeScan, statistical significance was established from a null distribution generated by running SweeD on SNP datasets simulated under the inferred demographic history for P. leucopus populations (Harris et al. 2016). SweeD does not inherently identify outlier regions. The CLR is computed using a selective sweep model on the observed data and then compared to a neutral model calibrated with a simulated background SFS. As before, we used 100 datasets with 100,000 SNPs each, simulated under the inferred neutral demographic history for white-footed mice in NYC. The CLR was calculated using SweeD for all simulated datasets. We identified outlier contigs if their CLR value was greater than any produced in neutral simulations. We also required outliers to fall within the top 0.01% of the CLR distribution for the observed SNPs.

Genotype-environment association tests for environmental selection

We used the GEA approach of LFMM: Latent Factor Mixed Models (Frichot et al. 2013) to associate our full SNP dataset with potential environmental selection pressures. LFMM examines associations between environmental and genetic variation while accounting for the neutral genetic background and structure between populations (Frichot et al. 2013). We tested three environmental variables associated with urbanization: 1) percent impervious surface (i.e. surfaces such as roads, rooftops, and other human infrastructure that do not absorb water calculated from USGS National Land Cover Data) within a 2 km (the approximate lifetime dispersal distance of white-footed mice) buffer around each sampling site’s GPS coordinate, 2) human density within a two-kilometer buffer around each sampling site’s GPS coordinate (calculated from US Census blocks), and 3) categorization of each site as urban, within NYC limits, or rural, undeveloped state park outside city limits (Coded as 0 or 1 in LFMM). Calculations were made in ArcGIS v10.1 (ESRI, Redlands, CA, USA) and were previously reported in Munshi-South et al. (2016). This previous analysis found that variables 1–2 were significantly associated with genome-wide variation in P. leucopus populations in the NYC metropolitan area. LFMM requires the user to define the number of latent factors, K, that describe population structure in the dataset. To identify the appropriate number of K latent factors, we performed a genetic PCA followed by a Tracy-Widom test to find the number of eigenvalues with P values ≤ 0.01 (Patterson et al. 2006; Frichot & François 2015). Based on this approach, we ran LFMM with default parameters except for K = 6, number of MCMC cycles = 100,000, and burn-in = 50,000. Using author recommendations, we calculated the median |z|- score from 10 replicate runs and then readjusted the p values. LFMM uses |z|- scores to report the probability of a SNP’s association with an environmental variable. Again, we controlled for FDR by using a q-value threshold of ≤ 0.1.

BayPass also includes an environmental analysis, so for comparison to LFMM we used the GEA test implemented in the BayPass AUX model that identifies genetic markers associated with population-specific covariates (Gautier 2015). For population covariates, we used the same environmental variables used in LFMM: site classification (i.e. urban or rural) as a binary covariate, human density, and impervious surface. We used the AUX model and again simulated pseudo-observed datasets (PODs) under the Inference Model to calibrate neutral distributions for Bayes Factors (BFs). BayPass uses BFs to associate SNPs with population specific covariates. SNPs with BF estimates greater than the 95% threshold determined from PODs were considered to be associated with population covariates. We further filtered associations by setting a cutoff for BF ≥ 20.

Functional annotation of candidate genes

We used the gene annotation pipeline in Blast2GO (Conesa et al. 2005; Götz et al. 2008) to identify sequences from the NCBI non-redundant protein database that were homologous to our outlier contigs identified above. We then retrieved associated gene ontology (GO) terms. Blast2GO retrieves GO terms associated with BLASTX hits and uses the KEGG database to describe biochemical pathways linking different enzymes (Ogata et al. 1999; Kanehisa et al. 2014). For downstream enrichment analyses, we also used the Ensembl gene annotation system (Aken et al. 2016) to find homologous Mus musculus genes for each P. leucopus contig. We further interpreted the outlier gene lists using g:Profiler (Reimand et al. 2016) to identify gene ontology terms enriched in our outlier gene list compared to the fully annotated Mus musculus genome. We used the g:Profiler webserver and identified enriched terms associated with outlier genes using default parameters and the Benjamini–Hochberg correction for multiple comparisons with an adjusted p-value < 0.05. Finally, we used REViGO to cluster GO terms and summarize them in a subset of terms based on semantic similarity measures (Supek et al. 2011).

RESULTS

Genetic diversity statistics

In total, we identified 154,770 SNPs for investigating patterns of genetic variation and performing tests of selection. Urban populations had a 50% decrease in nucleotide diversity compared to the rural populations, but mean Tajima’s D values for rural parks were consistently higher than for urban parks (Table 1). The average nucleotide diversity for all three rural populations was 0.224 ± 0.034 SE, while the average for urban populations was only 0.112 ± 0.019 SE. The average Tajima’s D within populations did not show substantial differences between populations (Table 1). For all populations, Tajima’s D was slightly positive. Average pairwise F_ST were the lowest between rural populations (0.018 ± 0.364 SE, CFP – HIP Table S1) and highest between urban populations (0.110 ± 0.520 SE, CP – FM Table S1). These F_ST values were similar to F_ST estimated using genome-wide SNP datasets (Munshi-South et al. 2016).

Table 1.

Summary population genomic statistics (mean ± standard error) for three urban and three rural populations of white-footed mice (Peromyscus leucopus) examined in this study.

Population	Nucleotide diversity (π)	Tajima’s D
Urban
CP	0.131 ±0.001	0.318 ±0.005
FM	0.112 ±0.001	0.301 ±0.006
NYBG	0.092 ±0.001	0.280 ±0.006
Rural
BHwwp	0.198 ±0.001	0.350 ±0.004
CFP	0.211 ±0.001	0.336 ±0.004
HIP	0.263 ±0.001	0.349 ±0.004

Outlier detection and environmental associations

We used BayeScan to identify 39 outlier SNPs exhibiting patterns of divergent selection between urban and rural populations (Fig. 2A, Table S2). There were no SNPs that exhibited signatures of balancing selection. F_ST values for outlier SNPs ranged from 0.21 – 0.33. BayeScan identified zero outlier SNPs in the simulated neutral dataset, and accordingly the 39 outlier SNPs from the observed data had q-values that were smaller than the most extreme values for the simulated data (q-value ≤ 0.6). We ran a similar test looking for patterns of divergence using BayPass. This analysis identified 56 SNPs that showed evidence of divergent selection (Table S2). We used PODs to estimate a null distribution and outlier SNPs had XtX values ≥ 8.35 (top 5% of the null distribution). There were 11 SNPs associated with diversifying selection in both the BayeScan and BayPass analyses.

Figure 2.

To identify signatures of selective sweeps, we used the CLR statistic implemented in SweeD. We found that CLR scores in the top 5% of the simulated distribution were generally 2–3X lower than values in the top 5% of the observed dataset. We ran SweeD on observed SNPs within individual contigs and identified outliers by filtering for a CLR score ≥ 3.53 (the maximum CLR from simulated data). We also chose regions that fell within the top 0.01% of the observed distribution (Fig. 2B); all outliers had CLR scores ≥ 4.97. SweeD identified regions with SFS patterns that fit a selective sweep model in 45 contigs within urban populations (Table S2). There was no overlap between outlier SNPs identified by SweeD and BayeScan/BayPass.

There were 131 SNPs associated with at least one of three environmental variables tested using LFMM (Fig. 3A, Table S2). There was zero overlap with outliers identified from BayeScan and only one SNP that overlapped between SweeD and LFMM. Three SNPs identified in BayPass as outliers showing signatures of diversifying selection were also associated with environmental covariates in LFMM (Table S2). All three SNPs were within genes associated with human density around sampling sites and one was associated with all three environmental covariates. In an analysis similar to LFMM, we used BayPass to also associate environmental variables, called population covariates, with allele frequencies. There were 143 SNPs associated with at least one of the three environmental covariates tested using BayPass (Table S2). From these 143, five overlapped with those showing signatures of divergent selection in BayPass and eleven overlapped with outliers in BayeScan.

Figure 3.

Across all tests, SNPs identified as outliers or associated with environmental variables were found in 381 contigs. We filtered this list down to a subset of 19 contigs (Table 2) that are the most likely candidates for directional selection due to urban selective pressures. We required these filtered candidate contigs to show a signature of diversifying selection between urban and rural populations (BayScan or BayPass) or a signature of a selective sweep (SweeD), and they had to be associated with an environmental variable (human density around parks, impervious surface) as identified in GEA tests (LFMM or BayPass).

Table 2.

Outlier loci (N = 19) identified in at least one test for selection (BayeScan, BayPass, or SweeD) and one GEA test (LFMM or BayPass_GEA). SNP shows the position in contig containing the outlier loci. Tests show which tests identified the SNP as an outlier: BPG = BayPass_GEA; BPD = BayPass_Diversifying; BS = BayeScan; SW = SweeD; LFMM = LFMM.

Contig	SNP	Ensemble Gene ID	Gene	Tests
27887-125	142	ENSMUSG00000029440	proteasome 26S subunit, non-ATPase, 9	BPG, BS
3135-709	210	ENSMUSG00000002320	transmembrane 9 superfamily member 1	BPG, BS
37015-34	35	ENSMUSG00000037287	tubulin folding cofactor E-like	BS, BPD
5754-511	168	ENSMUSG00000041161	OTU domain containing 3	BPG, BS
7280-442	336	ENSMUSG00000021287	X-ray repair complementing defective repair in CHC3	BPD, LFMM
2260-821	1387	ENSMUSG00000024045	A kinase (PRKA) anchor protein 8	BPG, BS
27691-127	162	NA	NA	BPG, BS
27707-127	567	ENSMUSG00000106907	autophagy related 2A	BPG, BS, BPD
3567-665	756	ENSMUSG00000001700	GRAM domain containing 3	BPG, BS, BPD
10099-359	1465	ENSMUSG00000024066	xanthine dehydrogenase	BPG, SW
12107-321	433	NA	NA	BPG, BS
124-2491	596	ENSMUSG00000064358	cytochrome c oxidase III	BPG, SW
12685-311	481	ENSMUSG00000035637	glyoxylate reductase/hydroxypyruvate reductase	BPG, BPD
17856-243	2695	ENSMUSG00000021091	serine peptidase inhibitor, clade A, member 3N	LFMM, SW
22102-206	245, 1029	ENSMUSG00000045868	GTPase, very large interferon inducible 1	BPG, BPD, LFMM
34737-52	125, 257	NA	NA	BPG, BPD
35973-42	17	ENSMUSG00000001173	oculocerebrorenal syndrome of Lowe	BPG, BS
38397-23	175	NA	NA	BPG, BS
8088-415	154	ENSMUSG00000002379	NADH dehydrogenase 1 alpha subcomplex 11	BPG, BPD, LFMM

Functional annotation

The full contig sequences containing outlier SNPs were obtained from the P. leucopus transcriptome (Harris et al. 2015) and used for functional annotation and analysis. We first tested the full set of 381 contigs identified by all outlier tests for overrepresented GO terms using g:Profiler. There were 260 overrepresented GO terms from the full outlier list (Table S3). We summarized this list using REViGO into 23 representative terms. The top representative term was lipid metabolism, followed by organic substance catabolism (Table S4). The list also includes lipid homeostasis and immune system processes.

We also looked for overrepresentation in the gene annotations associated with the filtered subset of 19 outliers and found related results (Table 3). There were 15 contigs homologous to known genes with functional annotation. Metabolic pathways were the most overrepresented group of gene ontology terms, and there were two biological functions associated with the most overrepresented GO terms from the full list. These included non-alcoholic fatty liver disease and regulation of protein kinase b (AKT) signaling.

Table 3.

Overrepresented gene ontology (GO) terms from g:Profiler (q-value < 0.05) for the 19 outlier loci from tests for both selection and GEA. Associated genes shows which ensemble gene homologs from Table 2 are associated with each overrepresented term.

Description	Annotation ID	P-value	Associated Genes
Negative regulation of protein kinase B signaling	GO:0051898	0.05	ENSMUSG00000024066, ENSMUSG00000041161
Cytochrome c oxidase, mitochondrial	CORUM:538	0.05	ENSMUSG00000064358
Ocrl-Cdc42 complex	CORUM:975	0.00914	ENSMUSG00000001173
Glyoxylate and dicarboxylate metabolism	KEGG:00630	0.0355	ENSMUSG00000035637
Homologous recombination	KEGG:03440	0.0512	ENSMUSG00000021287
Pyruvate metabolism	KEGG:00620	0.0464	ENSMUSG00000035637
Oxidative phosphorylation	KEGG:00190	0.00791	ENSMUSG00000002379, ENSMUSG00000064358
Alzheimer’s disease	KEGG:05010	0.0133	ENSMUSG00000002379, ENSMUSG00000064358
Huntington’s disease	KEGG:05016	0.0158	ENSMUSG00000002379, ENSMUSG00000064358
Non-alcoholic fatty liver disease (NAFLD)	KEGG:04932	0.0101	ENSMUSG00000002379, ENSMUSG00000064358
Glycine, serine and threonine metabolism	KEGG:00260	0.05	ENSMUSG00000035637
Metabolic pathways	KEGG:01100	0.0029	ENSMUSG00000001173, ENSMUSG00000002379, ENSMUSG00000024066, ENSMUSG00000035637, ENSMUSG00000064358
Parkinson’s disease	KEGG:05012	0.00899	ENSMUSG00000002379, ENSMUSG00000064358
Autophagy - other	KEGG:04136	0.0404	ENSMUSG00000106907
Caffeine metabolism	KEGG:00232	0.00742	ENSMUSG00000024066

DISCUSSION

In this study, we investigated patterns of divergent positive selection between urban and rural populations of P. leucopus, and identified significant associations between outlier SNPs and environmental variables relevant to urbanization. The majority of candidate loci were annotated with GO terms that are significantly associated with dietary metabolism, particularly breakdown of lipids and carbohydrates. We discuss what these findings mean for organisms inhabiting novel urban ecosystems, and more generally for understanding the ecological processes and time frame of local adaptation in changing environments.

Our previous study investigated non-synonymous polymorphisms in pooled transcriptome samples and we reported evidence for positive selection in genes dealing with metabolism, immunity, and methylation in NYC white-footed mice (Harris et al. 2013). This current study supports the phenotypic traits likely under selection in urban environments, identifying outlier genes that play major roles in metabolism, and to a lesser extent, immunity, but few outlier genes were identified in both the current and previous studies. The dataset analyzed here was much larger, included more sampling sites, and changed the inclusion criteria for outlier genes by using analyses that identify more recent signatures of selection, as opposed to longer-term evolutionary changes in non-synonymous substitutions. However, it is important to note that our study is still relatively small, including only six populations and eight individuals from each population. Increasing the number of individuals and sampling sites, especially including multiple cities as replicates, would likely greatly improve the associations found between environmental variables and allele frequencies (Lotterhos & Whitlock 2015). The latter approach may be unlikely, however, with each urban setting presenting a unique set of selective pressures leading to local adaptive responses, as shown with coat coloration in beach mice (Peromyscus polionotus) (Hoekstra et al. 2006) and climate related adaptation in the flowering plant (Arabidopsis halleri) (Rellstab et al. 2017). Despite potential issues with sample size, we did find two of the eleven previously identified candidate genes (Harris et al. 2013) to be direct matches to outliers in this current analysis (Serine protease inhibitor a3c and Solute carrier organic anion transporter 1A5), and two other genes were from the same gene families or involved in the same biological processes. One gene, an aldo-keto-reductase protein, is part of the same gene family as the aflatoxin reductase gene (Contig 10636-348) identified in this study. The aldo-keto reductase gene family comprises a large group of essential enzymes for metabolizing natural and foreign substances (Hyndman et al. 2003). The other is a cytochrome P450 (CYPA1A) gene involved in metabolism of drugs and lipids. Peromyscus directly express CYPA1A and Hsp90 (outlier from current SweeD analysis) when exposed to environmental toxins (Settachan 2001).

Population genomics summary statistics

Before performing outlier tests, we initially calculated per-site nucleotide diversity and Tajima’s D. The Tajima’s D statistic was calculated per contig for each population. We found nucleotide diversity to be lower in all urban population compared to rural populations, supporting previous work that found a negative association between genome-wide SNP diversity and urbanization. That study included the six populations studied here and an additional 18 populations distributed along an urban-to-rural gradient (Munshi-South et al. 2016). While loss of genetic variation will reduce evolutionary potential and decrease the probability of local adaptation, selection may still act if adequate variation is present and genetic drift is not too strong (Donihue & Lambert 2015; Munshi-South et al. 2016). Tajima’s D is often used to identify signatures of selection, comparing observed to expected heterozygosity. For all our populations, Tajima’s D skewed positive, possibly explained by balancing selection. While balancing selection has been found to maintain variation in immune loci in fragmented urban population of bobcats (Lynx rufus) (Serieys et al. 2015), it is difficult to distinguish whether demography or selection drives Tajima’s D values in many cases (MacManes & Eisen 2014). We have estimated the complex demographic history for P. leucopus populations in NYC (Harris et al. 2016), suggesting Tajima’s D may not be the best tool for identifying selection in this system. Outlier tests are more robust to demography and we explicitly accounted for the specific demographic history of P. leucopus in the null models used during analysis of our genome scan methods.

Signatures of selection in urban populations from genome-wide scans

Over the past decade, genome scans have become feasible methods to detect and disentangle neutral and adaptive evolutionary processes for non-model organisms (De Villemereuil et al. 2014; Hoban et al. 2016). One method, BayeScan (Foll & Gaggiotti 2008), calculates the posterior probability that a site is under the influence of selection by testing models with and without selection. While BayeScan is relatively robust to confounding demographic processes (Pérez-Figueroa et al. 2010; De Villemereuil et al. 2014), population bottlenecks, hierarchical structure, recent migration, or variable times to most-recent-common-ancestor (MRCA) between populations can artificially inflate F_ST values (Hermisson 2009; Lotterhos & Whitlock 2014) and may still impact BayeScan (Savolainen et al. 2013; Lotterhos & Whitlock 2014). We minimized false positives by incorporating population structure and a specific demographic history for P. leucopus in NYC directly into the null distribution of F_ST (Harris et al. 2016). We only included outliers if their posterior probability was greater than probabilities calculated from these simulations. The outliers from BayeScan comprised 0.024% of the total number of loci analyzed from our RNASeq dataset, and 0.036% of the total loci using BayPass. These percentages are in line with candidates uncovered from a similar study (0.05%) that looked at high and low altitude populations of the plant Senecio chrysanthemifolius (Chapman et al. 2013). Many studies find higher percentages of outlier loci using BayeScan; for example, 4.5% in the American pika across its range in British Columbia (Henry & Russello 2013), and 5.7% in Atlantic herring across their range (Limborg et al. 2012). Our lower overall percentage of outliers may be due to differences in species or datasets between studies (false positive rate, power, sampling, genome size and composition are all variables that influence numbers of SNPs), or alternatively because of relatively recent isolation or moderate to weak selection in urban populations.

SweeD, another genome scan approach, examines patterns within a population’s SFS rather than allelic differentiation between populations. The main footprint that selective sweeps leave on the SFS is an excess of low- and high-frequency variants (Nielsen 2005). The SweepFinder method (Nielsen et al. 2005), recently upgraded to the NGS compatible SweeD (Pavlidis et al. 2013), uses a CLR test based on the ratio between the likelihood of a neutral and selective sweep hypothesis. As above, the weakness of hitchhiking methods is the confounding influence certain demographic processes have on the SFS (Hermisson 2009). However, building a robustly inferred demographic history into the null model substantially reduces false positive rates (Pavlidis et al. 2013). We included the P. leucopus demographic history into our analysis, and found 0.019% of the sequenced loci to contain SFS patterns indicative of selective sweeps. This rate is in line with other studies that reported that 0.5% of regions in domesticated rice (Wang et al. 2014), 0.02% of loci in black cottonwood (Zhou et al. 2014), and 0.02% of the gorilla genome (McManus et al. 2014) show evidence of selective sweeps or hitchhiking.

Several studies have shown that identifying outliers with multiple tests and diverse theoretical approaches is the best way to reduce false positives in genome outlier analyses (Nielsen 2005; Grossman et al. 2010; Hohenlohe et al. 2010b). We required candidate genes to show a signature of diversifying selection or a signature of a selective sweep, and they had to be associated with an environmental variable. We found several outliers identified in both BayeScan and BayPass (Table S2), however, there was no overlap between BayeScan/BayPass and SweeD outliers. This discrepancy is likely due to the different selection scenarios underlying each test, i.e. divergent local selection versus population-wide positive selection in the form of selective sweeps (Hermisson 2009). F_ST based methods respond to allelic divergence relatively quickly, while models for selective sweeps typically require nearly-fixed derived alleles (Hohenlohe et al. 2010b). Given the recent history of urbanization in NYC, many selective sweeps may be ongoing or otherwise incomplete. Selection may also be acting on standing genetic variation in the form of soft sweeps (Hermisson & Pennings 2005) that are not readily identified by SweeD.

Environmental associations strengthen evidence of local adaptation to urbanization

GEA tests are a growing class of methods that identify loci that are associated with environmental factors (Joost et al. 2007; Coop et al. 2010; Frichot et al. 2013), and by accounting for underlying correlation structure of allele frequencies, may often be more powerful than traditional outlier tests (Savolainen et al. 2013). GEA tests come from the field of landscape genomics which incorporates tools from landscape genetics and population genomics to examine the effects of demography, migration, and selection, and ultimately identify local adaptation (Sork et al. 2013; Rellstab et al. 2015). Here we used LFMM (Frichot et al. 2013) and the AUX covariate model from BayPass on the full SNP dataset with environmental metrics of urbanization. LFMM performs better than other methods in the presence of hierarchical structure and when polygenic selection is acting on many loci with small effect (De Villemereuil et al. 2014). Hierarchical structure in our dataset includes urban and rural differentiation (Harris et al. 2015; Harris et al. 2016), patterns of geographic structure between mainland mice and Long Island, NY (Harris et al. 2016), and population structure between individual urban parks (Munshi-South & Kharchenko 2010). Simulations also suggest that LFMM is superior when sample size is less than 10 individuals per population, there is no pattern of IBD, and the study compares environmentally divergent habitats (Lotterhos & Whitlock 2015). We sampled eight white-footed mice per population, found no evidence of IBD (Munshi-South et al. 2016), and sampled environmentally divergent rural and urban locations.

Using GEA tests implemented in BayPass and LFMM, we found that 17 (12 %) and 4 (2.8 %) outliers, respectively, were significantly associated with one or more urbanization variables. These results are lower than other studies combining genome scans and GEA tests. Limborg et al. (2012) found 62.5% of the outliers identified in BayeScan were correlated with temperature or salinity in Atlantic herring, and 26.3% of genome scan outliers were associated with temperature or latitude in a tree species (De Kort et al. 2014). The lower overlap found in our study is likely due to the difficult nature in quantifying urbanization. Percent impervious surface, human population density, or binary classification as urban versus rural may not capture the specific, causative selection pressures acting on white-footed mouse populations (See Table S5 for environmental data). We used these metrics as general proxies for changing ecological processes in urbanized habitats. The percent of impervious surface around a park is likely representative of habitat fragmentation, as urban infrastructure changes the net primary productivity due to increasing percentages of impervious surface or artificial landscapes, parks and yards (Shochat et al. 2006). This fragmentation then leads to changing species interactions as migration is impeded or organisms are forced into smaller areas (Shochat et al. 2006). The percent human density surrounding an urban park can serve as a proxy for the multitude of ecological changes humans impose on their surrounding environment. Urbanization and increasing human density change the types and availability of resources in the altered habitat (McKinney 2002; Sih et al. 2011). Finally, classifying our sites as urban or rural can generally capture the main differences in urban and natural sites. For example, pollution is a major consequence of urbanization (Donihue & Lambert 2015), and urban areas often include increased chemical, noise, or light pollution (Sih et al. 2011).

Between divergent allele frequencies, a skewed SFS, environmental associations, and overrepresented GO terms, we find several overlapping lines of evidence that support rapid divergent selection in white-footed mice. Our results support the growing body of evidence (Donihue & Lambert 2015) that finds urbanization directly impacts the ecology and evolution of species. However, to fully support the hypothesis that organisms adapt to urban habitats, it is still necessary to link genetic changes to measurable phenotypic differences and measure direct fitness benefits. Past urban evolutionary studies often focus solely on phenotypic (Yeh 2004; Partecke et al. 2006; Cheptou et al. 2008; Thompson et al. 2016) or genetic (Wandeler et al. 2003; Noël & Lapointe 2010; Mueller et al. 2013; Lourenco et al. 2017) differences between populations in and outside of cities. However, researchers are beginning to examine both the genotype and phenotype in parallel instances of urban evolution (Whitehead et al. 2010; Wirgin et al. 2011; Hof et al. 2016), which is key to understanding how urbanization affects the evolution of species. In the future, the gene annotations for our predicted outlier genes can help determine which phenotypic traits to measure in urban P. leucopus populations.

Functional roles of candidate genes: quality of urban diet?

The model rodents Mus musculus, Rattus norvegicus, and Cricetulus griseus all have deeply sequenced, assembled and annotated reference genomes. These resources allowed us to annotate 89.5% of outlier loci with high quality functional information. Urban P. leucopus exhibited signatures of positive selection in genes with GO terms overrepresented for organismal metabolic processes, specifically digestion and metabolism of lipids and carbohydrates.

Mitochondrial genes identified as outliers (Table S2) were largely responsible for the overrepresentation of metabolic process. While we can only speculate until further physiological studies are conducted, our evidence suggests that the evolution of mitochondrial and metabolic processes has been important to the success of P. leucopus living in NYC’s urban forests. Mitochondrial genes have often been used to describe neutral population variation, but researchers have found ample evidence of selection acting on the mitochondrial genome (Oliveira et al. 2008; Balloux 2010). For example, specific mitochondrial haplotypes are associated with more efficient thermogenesis and higher fitness in over-wintering shrews (Fontanillas et al. 2005). Pergams & Lacy (2007) found complete mitochondrial haplotype replacement in contemporary P. leucopus in Chicago compared to haplotypes sequenced from museum skins collected before urbanization. The agent of selection is not clear, but Munshi-South and Nagy (2014) also identified signatures of selection (or alternatively population expansion) in mitochondrial D-loop haplotypes from contemporary P. leucopus in NYC. Many mitochondrial functions are affected by the same environmental variables that change in response to urbanization, such as temperature (Balloux 2010), reduced migration (Lankau & Strauss 2011; Munshi-South 2012), or resource availability (Burcelin et al. 2002).

Urban P. leucopus may experience different energy budgets, physiological stressors or diets compared to rural counterparts. We found a substantial number of candidate genes with functions related to the metabolism and transport of lipids and carbohydrates, and the most common overrepresented GO terms involved lipid metabolism and homeostasis (Table S4). In the full outlier analysis, two genes are particularly interesting as targets of diet-mediated selection. The first gene, FADS1, is a fatty acid desaturase important for the biosynthesis of omega-3 and -6 fatty acids (long-chain polyunsaturated fatty acids, LCPUFA) from plant sources. Recent evidence suggests that the FADS gene family has been an important target of selection in humans during the transition from hunter-gather to agricultural societies (Ye et al. 2017). Alleles linked to upregulated biosynthesis of LCPUFAs (naturally low in plant based diets) increased in frequency after the Neolithic Revolution (Ye et al. 2017). We aligned our homologous FADS1 contig with human transcripts to identify whether P. leucopus had any relevant alleles, but our sequenced populations did not contain SNPs at any relevant loci. The full list of outliers also contained APOB-100, which is the primary apolipoprotein that binds and transports lipids, including both forms of cholesterol (HDL and LDL).

When we investigated only candidate genes that were identified by both an outlier test and GEA test, we found similar patterns suggesting P. leucopus in urban environments may be adapting to novel food resources. These genes were strongly correlated with environmental measures of urbanization, with clearly divergent allele frequencies between urban and rural sites (Fig. 3B), suggesting that selection is acting on standing genetic variation in urban environments. The most significant overrepresented GO term involved regulation of protein kinase B (AKT). AKT is a key molecule in the insulin signaling pathway, important for promoting glucose storage and regulating glucose in the bloodstream between fed and fasting states (Boucher et al. 2014). Glycine metabolism was also overrepresented; increased amounts of glycine may be important for regulating high-fat, high-sugar diets by decreasing concentrations of free fatty acids and triglycerides (Wang et al. 2013). Finally, our candidate list contained genes significantly associated with non-alcoholic fatty liver disease (NAFLD). NAFLD is a major hallmark of obesity and diabetes and can be induced through increased uptake of dietary fatty acids (Fabbrini et al. 2010).

These candidate genes suggest that white-footed mice in isolated urban parks may be evolving in response to food resource differences between urban and rural habitats. This finding is corroborated by recent evidence that urban white-footed mice in NYC have shorter upper and lower tooth rows than rural mice (Yu et al. 2017). Lower quality food in the diet often requires increased chewing and is accompanied with larger occlusal surfaces, and subsequently, longer toothrows (Ungar 2010). One prediction is that urban P. leucopus consume a diet with a substantially higher fat content than diets of rural populations. The typical diet of P. leucopus across its range consists of arthropods, fruits, nuts, various green vegetation, and fungus (Wolff et al. 1985). Given that white-footed mice are opportunistic generalists, many different food resources could differ between urban and rural habitats. Urbanization in NYC has produced relatively small green patches that are surrounded by a dense urban matrix, and P. leucopus in NYC may successfully take advantage of invasive plant species, different arthropod communities, or increased human food waste in and around their urban habitats. Local adaptation in urban populations may allow these mice to more efficiently metabolize different types or amounts of lipids and carbohydrates, although field studies are needed to examine the link between these genetic changes and diet in NYC.