Introduction

Climate change is occurring at rates likely to exceed the capacity of many populations to adapt via contemporary evolution, casting doubt on their persistence (Anderson et al. 2012). Predicting ‘winners and losers’ in this race remains challenging due to uncertainties about how to characterize the adaptive potential of populations, the effects of historic and contemporary gene flow and natural selection on the pace of local evolution, and the appropriate metrics for prioritizing populations for conservation (Bay et al. 2018; Coates et al. 2018). A growing body of literature aims to define populations’ adaptive capacity, or the ability to cope with environmental change through phenotypic plasticity and evolution (Bay et al. 2017) and to use this information to help inform conservation initiatives. Historically, and in the absence of genetic data, biologists often rely on phenotypic traits to delineate populations for conservation, especially those traits thought to be under genetic control, influential to fitness, or diagnostic of population identity (Haig and Winker 2010). The increasing application of modern genomic methods for systematics, however, provides the opportunity to evaluate adaptive and neutral evolutionary processes, offering additional support for conserving population-level variation (e.g., Oh et al. 2019).

Although these approaches are promising, from a taxonomic perspective, there are concerns that the increased resolution from genomic data may lead to over-splitting if all genetically distinctive populations are classified as full species (Coates et al. 2018; Winker 2021). For birds, most phenotypically distinctive populations have historically been described and formally named as taxonomic subspecies, and in turn these subspecies have often become populations targeted for increased conservation efforts and legal protection (Barrowclough et al. 2016). Some taxonomists, however, have criticized the prioritization of subspecies in conservation, particularly when the subspecies’ classification is based on delineations supported by small numbers of neutral genetic markers, or on morphological traits that are clinal across space (Zink 2004, 2010; Zink and Barrowclough 2008). Such debates arise via the worthy desire to discretize biological diversity, often as a requirement of legal proceedings and legislative policies, versus the contrasting view of speciation and differentiation as an ongoing continuum, a perspective increasingly supported by genomic data (Henderson et al. 2020).

This debate is particularly well developed with respect to a San Francisco Bay complex of five resident song sparrow subspecies: Melospiza melodia heermanni, M. m. maxillaris, M. m. samuelis, M. m. pusillula, and M. m. gouldii (Fig. 1a,c). Song sparrows are widely distributed across North America and display phenotypic variation among their 25 named and 52 described subspecies. The San Francisco Bay region contains the highest concentration of individual subspecies, with five recognized subspecies occupying distinct habitats within a 100 × 70 km region that spans a selective gradient of salt and freshwater habitats. Maxillaris, samuelis, and pusillula are particularly notable for their year-round residency restricted to separate but geographically proximal salt marshes; they occupy a challenging niche that only 25 other vertebrate species have been able to successfully colonize, of which nearly all are of conservation concern due to habitat loss (Greenberg et al. 2006). To this end, subspecies delineations in this system are supported by evidence suggesting rapid adaptation to saline environments (Basham and Mewaldt 1987), differences in coloration (Miller 1956), and genetic isolation (Ferrell 1966), all amounting to substantial evidence of parapatric differentiation (Mayr 1942). In contrast, neither microsatellite (Chan and Arcese 2002) nor mitochondrial DNA (Fry and Zink 1998) data are coincident with these phenotypes (Patten and Pruett 2009). For example, Zink (2004, 2010) concluded that neither subspecies delineations nor neutral genetic markers offer evidence of evolutionary significance or conservation value for these localized populations of sparrows. Most recently, reduced representation genome sequencing provided increased resolution, with subtle separation of the subspecies at microgeographic scales (Mikles et al. 2020). Though the underlying mechanisms of such genomic divergence remain unclear, differentiation is not explained by isolation by distance (Chan and Arcese 2002; Mikles et al. 2020). Here, we use the San Francisco Bay song sparrows as a case study to ask if their subspecies delineations reflect evolutionary processes responsible for the rapid diversification of these populations over the last 10,000 years (Chan and Arcese 2002). If so, then these subspecies classifications would have utility in indicating their evolutionary distinctiveness, and this would in turn imply that the subspecies level of classification has potential merit for recognizing conservation units in other avian taxa.

Fig. 1
figure 1

(A). San Francisco Bay song sparrow ranges digitized. Black dots indicate the sampling locations from the five representative populations and the colors represent the ranges for each subspecies in the Bay (gouldii in red, heermanni in green, samuelis in blue, maxillaris in yellow, and pusillula in orange. (B). Variable importance output ordered by their importance estimated by the RF model. Bio4 = Temperature Seasonality, Bio11 = Mean Temperature of the Coldest Quarter, Bio1 = Annual Mean Temperature, salt = Salinity, bio12 = Annual Precipitation, Bio16 = Precipitation of Wettest Quarter, Bio19 = Precipitation of Coldest Quarter, Bio18 = Precipitation of Warmest Quarter, Bio9 = Mean Temperature of Driest Quarter, Bio10 = Mean Temperature of Warmest Quarter, Bio15 = Precipitation Seasonality, Bio17 = Precipitation of Driest Quarter. (C). Song sparrow subspecies illustrations by Jillian Ditner demonstrate subtle morphological differences among the subspecies. (D). Variation in the top 5 most important variables from RF model by subspecies. Variables were normalized to bring values to range from 0–1. Dark horizontal lines represent the median, colored boxes show the interquartile range, whiskers indicate the 5th – 95th percentile, and dots represent outliers

Full size image

Methods

Whole genome re-sequencing and variant discovery

We sequenced the genomes of 43 song sparrows representing five subspecies from the San Francisco Bay area (Fig. 1a; Table S1). All birds were sampled during the breeding season (March to May, 1999) by Y. Chan and P. Arcese; adults were captured in mist nets, measured and blood sampled, and released (see Chan and Arcese 2002). We extracted genomic DNA using the DNeasy blood and tissue kit (Qiagen, CA, USA) and quantified DNA concentrations using the Qubit dsDNA High Sensitivity Assay Kit (Life Technologies). Using 200 ng of DNA from each sample, we prepared individually barcoded libraries with a 550 bp insert size following the protocol for the TruSeq Nano DNA Library Prep kit (Illumina, California, USA). We then sequenced libraries for heermanni, maxillaris, and samuelis on a single Illumina NextSeq lane at the Cornell Institute for Biotechnology core facility. We obtained raw sequences for gouldii and pusillula from Walsh et al. (2019a).

We assessed library quality using FastQC version 0.11.8. We used AdapterRemoval version 2.1.1 for sequence trimming, adapter removal, and quality filtering, requiring a minimum Phred quality score of 20 and merged overlapping paired-end reads. We aligned filtered reads to the Song Sparrow reference genome (Feng et al. 2020) using the default settings in BWA 0.7.4 (Li and Durbin 2009) and obtained alignment statistics from Qualimap version 2.2.1 (Okonechnikov et al. 2016). We removed three samples for > 50% missing data, and one for 35% relatedness to another individual (selecting to keep the individual with higher quality mapping statistics). We used Samtools version 1.9 (Li et al. 2009) to convert all resulting BAM files to SAM files and to sort and index files, and Picard Tools version 2.19.2 to add mark duplicates. We used the Haplotype Caller module in GATK version 3.8.1 (McKenna et al. 2010) for single nucleotide polymorphism (SNP) discovery and genotyping and used the following filtering parameters to remove variants: QD < 2, FS > 60.0, MQ < 30.0, and ReadPosRankSum < -8.0. We additionally filtered out variants that were not biallelic, had minor allele frequencies less than 5%, mean coverage less than 2X or more than 50X, and more than 20% missing data. This resulted in a total of 1,630,425 SNPs across all five subspecies.

Characterizing subspecific niche variation

We used existing range maps, bioclimatic variables (WorldClim V2; Fick and Hijmans 2017), and Random Forest (RF; Breiman 2001) to characterize environmental niches for each subspecies to assess possible patterns of ecological divergence in this system (Elith and Leathwick 2009). Briefly, we first georeferenced range maps in Chan and Arcese (2002), Patten et al. (2004), and Patten and Pruett (2009) using georeferencing and editor tools in ArcGIS (version 10.7.1; ESRI 2019) to obtain a contiguous map of the range of each subspecies (Fig. 1a). We note here that our digitized range maps are consistent with maps dating as far back as 1948 (Marshall 1948a) and are an accurate reflection of population ranges through time. To characterize variation in their niches we used a priori knowledge on the effect of climatic conditions and events on song sparrow demography to identify 11 candidate variables: Annual Mean Temperature (bio1), Temperature Seasonality (bio4), Mean Temperature of Driest Quarter (bio9), Mean Temperature of Warmest Quarter (bio10), Mean Temperature of the Coldest Quarter (bio11), Annual Precipitation (bio12), Precipitation Seasonality (bio15), Precipitation of Wettest Quarter (bio16), Precipitation of Driest Quarter (bio17), Precipitation of Warmest Quarter (bio18), Precipitation of Coldest Quarter (bio19) (Table S2) extracted at 30 arc-second spatial resolution from WorldClim, spanning the period of 1970–2000. As an additional candidate variable, we estimated salinity as a proxy using Euclidean distance to the edge of the nearest saline wetland using the spatial analyst tool in ArcGIS and the Areas of Conservation Emphasis Saline Wetlands geospatial data layer, which represents wetland area and location at the HUC12 watershed level (California Department of Fish and Wildlife, 2019). This salinity proxy was determined assuming there is a plateau, and thus was calculated as the Euclidean distance squared, resulting in the further distances being much larger and therefore less affected by salt. Values obtained via our Euclidean distance measurements are an accurate reflection of salinity measurements in the Bay (https://pubs.usgs.gov/of/2017/1022/ofr20171022.pdf), and thus offer a reliable proxy for salinity at our sampling locations.

We used RF to classify and delineate subspecies ranges by their environmental conditions. Within the RF algorithm, trees are produced using a bootstrapped sample of ‘bagged’ data that comprise ~ 64% of all observations, and they are tested against the remaining ‘out-of-bag’ (OOB) data to estimate prediction error (OOB error) as the percentage of incorrectly classified observations. We carried out the tuning and fit of RF in R version 4.0.2 (R Core Team 2020), using the “Caret” and “randomForest” packages (Liaw and Wiener 2002; Kuhn 2020). Our model was trained using five repetitions of a 10-fold cross validation scheme, wherein we randomly split the data into 50% training (n = 30,000) and 50% evaluation data (n = 30,000) to avoid overfitting (Kuhn and Johnson 2013). These data were generated by randomly selecting 60,000 pixels across our study area and associated with subspecies ranges (Fig. 1a). This process generated a final model trained with 11 predictor variables with the best hyperparameter values (mtry = 2; ntree = 500) according to the RMSE parameter (Kuhn and Johnson 2013). To reduce prediction bias, we fit a balanced RF model by growing each tree from random samples of the data with an equal number from each class. We adopted the Gini index to evaluate variable importance, wherein a higher Gini importance indicates a variables’ importance in maintaining predictive power in the RF model. We then evaluated model performance using unseen evaluation data for accuracy, sensitivity, specificity, and Kappa (Kuhn and Johnson 2013). To assess whether the observed subspecies niches were statistically different among the song sparrows, we used an analysis of variance (ANOVA) on the top five most important variables from the final RF model. Variables in this analysis were normalized to bring values to range from 0 to 1 to assess effect size.

Neutral genomic population structure & subspecies delineation

To characterize patterns of genetic structure among subspecies, we ran Admixture version 1.2.3 (Alexander et al. 2009) using a filtered data set (4,961 SNPs) that contained no missing data, was pruned to avoid linkage using the script ldPruning.sh (https://github.com/speciationgenomics/scripts/blob/master/ldPruning.sh), and had putatively adaptive SNPs removed (see below). For this analysis, we investigated five population clusters, using the default settings. For all other analyses of population structure, we assessed patterns and genomic diversity based on the full data set (1,630,425 SNPs) and a putatively neutral SNP data set. To target putatively neutral SNPs, we excluded SNPs that were mapped to exons or intervals within 25 kb of exons (see Kawakami et al. 2014), resulting in 308,973 SNPs. We performed a principal component analysis (PCA) on the full and neutral data sets using the “SNPRelate” package in R (Zheng et al. 2012). For both data sets, we additionally characterized genome-wide patterns of divergence between subspecies by calculating pairwise FST values for each comparison using VCFtools (Danecek et al. 2011). We calculated FST for 25 kb windows across our scaffolds and for individual SNPs, dropping windows with fewer than 10 SNPs. Using pairwise FST estimates for the full data set, we tested for isolation by distance using a Mantel test in R. We quantified genetic diversity by estimating individual heterozygosity and nucleotide diversity in 25 kb windows using VCFtools; for these calculations, we removed all missing data from both datasets, as estimates can be biased by missing data (Schmidt et al. 2021). For the full and neutral data sets, we counted the number of private alleles within each subspecies using bcftools (Li et al. 2009).

Genotype-environment associations

To assess whether the observed genomic differentiation among song sparrows in the San Francisco Bay area was a result of ecological divergence, we scanned for SNPs associated with five environmental variables identified by RF as most important in delineating ecological niches of the subspecies. These variables included temperature seasonality, mean temperature of the coldest quarter, annual mean temperature, salinity, and annual precipitation. We tested for Genotype-Environment Associations (GEA) through a combination of multivariate and univariate approaches: we ran a redundancy analysis (RDA; multivariate) using the rda function in the R package Vegan 2.4-5 (Oksanen et al. 2017), and a latent factor mixed model (LFMM; univariate) using the lfmm function of the LFMM package in R. Both methods are robust to a range of underlying demographic processes and sampling designs (Rellstab et al. 2015; Forester et al. 2018), while providing a balance between error rates and detection power (Carvalho et al. 2020). For both methods, we imputed missing genotypes for the full data set by using the most common genotype at each SNP across all individuals. Because temperature variables were correlated (pairwise Pearson correlation coefficients > 0.7), we ran a PCA on all temperature variables using the prcomp function in R and used the first and second principal components as predictors in the GEA analyses (Frichot et al. 2013). We used results from ADMIXTURE to define the number of latent factors used as K = 3 (Supplemental Material, Figure S1).

RDAs can offer a robust approach to detecting correlations between genotype and environmental data, particularly compared to other differentiation-based outlier scans (Rellstab et al. 2015; Forester et al. 2018). We used an RDA to test for multilocus signatures of selection for multiple environmental variables and evaluated the significance of the RDA using an analysis of variance (ANOVA) with 999 permutations. Loci with a loading ± 4 SD from the mean loading on the first three constrained ordination axes were considered candidates under environmental selection following Forester et al. (2018). We used a Pearson’s correlation (r) to identify environmental variables exhibiting the strongest association with each candidate locus.

For the LFMM analysis, we conducted 5 independent runs of 10,000 iterations and a 5,000 iteration burn-in. We corrected association p-values based on empirical genomic inflation factors (𝝀) to control for false discoveries (François et al. 2016). To do this, we inspected the distribution of p-values to ensure that they followed a normal distribution following François et al. (2016). For all variables, we set 𝝀 = 0.1. When generating a list of candidate SNPs, we used the Benjamini-Hochberg algorithm (Benjamini and Hochberg 1995) with a maximum false discovery rate of 10−5.

Evolutionary distinctiveness and identification of conservation units

Based on the GEA methods above, we identified shared candidate SNPs that correlated with environmental variables, which we refer to from here on as a putatively adaptive SNP data set. To provide a measure of adaptive diversity we calculated pairwise FST and individual heterozygosity with the putatively adaptive SNP data set. To prioritize groups for conservation, we calculated a measure of population distinctiveness for each subspecies. To do this, we calculated the Shapley metric (SH; Volkmann et al. 2014), which can be calculated from unrooted trees. FST estimates from pairwise comparisons of all five subspecies were used to build a NeighborNet network using the neighborNet function in the R package “phangorn” (Schliep 2011; Hudson and Bryant 2006). This network was then used to estimate the genetic contribution of individual tips (Volkmann et al. 2014), with higher SH values indicating higher priority for management.

Results

Whole genome re-sequencing yielded a mean of 16,043,533 reads per individual with the following sample sizes per subspecies: gouldii (n = 10), heermanni (n = 8), samuelis (n = 6), pusillula (n = 9), and maxillaris (n = 6). The mean alignment rate was 97.5%, the mean coverage was 2.60X, and the mean missing data was 12% (Table S1).

Subspecific niche variation

Climate varied predictably across the five subspecies ranges. The top five most influential variables in delineating subspecies niches were temperature seasonality (bio4), mean temperature of the coldest quarter (bio11), annual mean temperature (bio1), annual precipitation (bio12), and distance to saltwater (Fig. 1b). The influence of the top five climate predictors varied significantly among subspecies ranges (Fig. 1d; p < 0.001), except coldest quarter (bio11) and annual mean temperature (bio1) did not differ between the ranges of gouldii (normalized mean ± SD: 0.47 ± 0.17) and maxillaris (0.47 ± 0.05; p > 0.99), and gouldii (0.46 ± 0.19) and samuelis (0.48 ± 0.05; p > 0. 05), respectively. Distance to saltwater over the ranges of the marsh and upland subspecies varied as predicted, given their known ecotypes. Distance to saltwater also varied between the two upland subspecies (gouldii: 0.13 ± 0.18; heermanni: 0.22 ± 0.25; p < 0.001).

The final RF model displayed high accuracy (97.92% ± 0.21), sensitivity (91.81% ± 0.48), and specificity (98.49% ± 0.19), and a mean error rate of 2.27% (± 0.08) and Kappa of 96.75% (± 0.32). Classification error differed among subspecies such that upland subspecies had the lower error rates (gouldii = 2.25%; heermanni = 0.96%) compared to marsh subspecies (maxillaris = 8.58%; pusillula = 15.14%; samuelis = 16.72%).

Neutral genomic Population structure & subspecies delineation

We observed subtle differentiation among the five subspecies. In the full data set, results from Admixture supported K = 1 as the best supported cluster, yet we detected observable structuring up to K = 3. Under this model, clusters corresponded to (1) pusillula, (2) gouldii, and (3) heermanni, maxillaris, and samuelis (Figure S1;a). Based on ~ 1.6 million SNPs, we observed some clustering by subspecies on a PCA (Figure S1;b), with the three salt marsh subspecies separating along axis one (3.92% of variation explained). The most prominent clustering in this PCA was separation of pusillula and samuelis from the rest of the subspecies. This pattern of separation along PC axes was still present but less pronounced for the neutral data set (Figure S2). Genome-wide FST estimates further suggest moderate levels of divergence among San Francisco Bay area song sparrows (Table S3, S4). Mean FST estimates based on the full data set ranged from 0.011 (maxillaris vs. heermanni) to 0.047 (pusillula vs. samuelis; Table S3). Per SNP FST estimates ranged from 0 to 1 in each of the 10 pairwise comparisons. There was no significant difference between FST estimates based on the full data versus FST estimates based on the neutral data set (Paired Sample T-test; df = 9, t = -1.67, p = 0.12). We found no significant correlation between genetic and geographic distance (p = 0.129; Figure S5).

We observed slight differences in heterozygosity across the five subspecies (Figure S3), with pusillula and gouldii populations exhibiting the lowest mean heterozygosity (0.33 and 0.35, respectively). Mean heterozygosity was highest in samuelis (0.46). We observed a significant difference between individual heterozygosity estimates based on the full versus neutral data set (Paired Sample T-test; df = 38, t = -14.7, p < 2.2e-16). Patterns of nucleotide diversity were notably similar among the five subspecies (Figure S4). Based on the full data set, the number of private alleles observed in each subspecies was: 13,247 (0.82%) in gouldii, 11,258 (0.70%) in heermanni, 4,279 (0.29%) in maxillaris, 16,140 (0.98%) in pusillula, and 10,156 (0.62%) in samuelis. The number of private alleles observed in the neutral data set was proportionally comparable: 0.81% in gouldii, 0.88% in heermanni, 0.37% in maxillaris, 1% in pusillula, and 0.78% in samuelis.

Genotype-environment associations

For the RDA, the first three components explained 36.15%, 27%, and 20.4% of the variation, respectively and the full model was significant (p = 0.027). Temperature PC1 showed significant variation with song sparrow genotypes (p = 0.014) and captured approximately 90% of the variation driven by all three temperature variables. Annual precipitation (bio12; p = 0.14), temperature PC2 (p = 0.84), and salinity (p = 0.1) did not show significant variation with genotype. The first two axes of the RDA largely separated pusillula and gouldii from the other three subspecies (Fig. 2a). RDA1 appeared to associate more with salinity and RDA2 associated with the remaining environmental variables. We saw distinct clustering of individuals by subspecies along axes two and three of the RDA, with axis three separating salt marsh populations from upland populations (Fig. 2b). RDA3 appeared to be associated with all the variables, with salinity and annual precipitation being negatively correlated with temperature. Based on our cutoff of ± 4 SD, we identified 171 candidate SNPs that correlated with environmental variables. These included 143 SNPs associated with annual precipitation, 25 SNPs associated with temperature PC1, 1 SNP associated with temperature PC2, and 2 SNPs associated with salinity. Correlations between these candidate SNPs and their most strongly associated environmental variable were moderate, averaging 0.63 (r range = 0.13–0.74).

Fig. 2
figure 2

Subspecies delineation and evolutionary distinctiveness in San Francisco Bay song sparrows. (A) RDA axes 1 vs. 2 and (B) RDA axes 1 vs. 3 support increased divergence among subspecies driven by environmental factors. (C) NeighborNet depicting the relationship among song sparrows. Circle colors correspond to subspecies. Values associated with each tip represent SH metrics for each subspecies, indicating prioritization for management. Legend colors correspond to all panels in this figure

Full size image

LFMM identified substantially more candidate loci than the RDA. LFMM identified allele frequencies of 11,451 SNPs as significantly associated with environmental variables (282 SNPs for salinity; 8,172 SNPs for annual precipitation; 39 SNPs for temperature PC1; 2,958 SNPs for temperature PC2). Given the discrepancy in the number of outliers identified by LFMM compared to our other methods, we only retained those outliers that were identified by both the RDA and LFMM. For the downstream identification of genes linked to putative regions under selection, we retained 144 SNPs that were identified as outliers by both RDA and LFMM (Table S5).

Evolutionary distinctiveness and identification of conservation units

Mean FST estimates based on the 144 candidate SNPs ranged from 0 (samuelis vs. heermanni and gouldii vs. heermanni) to 0.058 (pusillula vs. samuelis; Table S6). Pairwise FST estimates did not differ significantly between putatively adaptive SNPs compared to estimates based on the full data set (Paired Sample T-test; df = 9, t = -2.56, p = 0.8). The overall pattern of heterozygosity among subspecies based on outlier SNPs was the same as that based on the full data set, with gouldii and pusillula having the lowest per-individual heterozygosity estimates. We observed a significant difference between individual heterozygosity estimates based on the full versus adaptive data set (Paired Sample T-test; df = 38, t = 2.66, p = 0.01). The NeighborNet network for subspecies in the San Francisco Bay is non-tree like, and the close placement of groups to each other on the network is as expected given their recent divergence times (Fig. 2c). We do see the placement of pusillula at a more isolated tip in the network, which is consistent with our other metrics of population structure. The pusillula subspecies had the highest SH rank, as expected based on the network (0.029), followed by maxillaris (0.01). The other subspecies largely clustered together in the network and had lower, and comparable, SH rankings: 0.008 for samuelis, 0.006 for heermanni, and 0.007 for gouldii.

Discussion

An extraordinary amount of attention has been dedicated to the study of the San Francisco Bay song sparrows (Grinnell 1909; Huxley 1942; Grinnell and Miller 1944, 1956, Marshall 1948a,b, Johnston 1956a,b, Mayr 1963, Chan and Arcese 2002, 2003, Mikles et al. 2020). Notable for their high concentration of morphologically distinct subspecies within a small geographic area, these song sparrows present a tractable system for investigating replicated colonization of marsh environments across a habitat gradient variable in salinity, temperature, and precipitation, and one now drastically altered by anthropogenic influence. By combining genomic data with random forest niche modeling, we identified evidence of local selection and putative ecological divergence over a fine spatial scale in the song sparrow subspecies of the San Francisco Bay. We posit that ecological variables linked to microgeographic habitat variation in the Bay are primarily responsible for the rapid radiation of these subspecies. Our work further suggests that selection for adaptive phenotypes, rather than neutral processes linked to drift or divergence time alone, is the primary driver of diversification in this system, however we discuss these conclusions within the context of alternative mechanisms in greater detail below. We conclude that these song sparrows offer lessons for the application of how genomic data can be applied to the characterization and conservation of local genetic diversity.

Evolution and conservation in the San Francisco Bay ecoregion

Saltwater marshes are inherently challenging environments that require specialized adaptations in the vertebrate species that have colonized them (Greenberg et al. 2006). The fragmented and patchy distribution of salt marsh habitats creates further challenges when prioritizing conservation efforts among small populations subject to rapid environmental change. Despite being the largest estuary on the west coast of North America, 90% of the San Francisco Bay marsh habitats have been converted to human use (San Francisco Bay Estuary Project 1991, Takekawa et al. 2006) and those remaining are at risk due to sea level rise in the next century (Thorne et al. 2018). Given the associated conservation challenges, active management of tidal marsh endemics is warranted. By identifying environmental drivers of locally adapted populations in these ecosystems, we can better understand how to preserve the full range of endemic phenotypes represented in the region.

Fine-scale mapping of habitats which appear to be homogenous, such as salt marshes, supported our hypothesis that the ranges of song sparrow subspecies in the San Francisco Bay area varied in microclimate. While we expected variation to be high between salt marsh and upland subspecies ranges, we also observed variation within both salt marsh and upland habitats, consistent with the hypothesis that environmental heterogeneity can facilitate local adaptation at fine spatial scales (Miller 1956; Mayr 1963; Ferrell 1966). Temperatures diverged most between the niches of the two upland subspecies, with gouldii experiencing the coolest conditions on the coast, and heermanni experiencing the warmest in inland habitats. Additionally, salinity varies greatly between upland and salt marsh subspecies but also within the estuary, with the South San Francisco Bay roughly 33 times saltier than the Suisun Bay, which receives large freshwater inputs from the Sacramento and San Joaquin Rivers (Schraga and Cloern 2017). Given ample evidence of rapid adaptation to saline environments in other species of new world sparrows (Walsh et al. 2019a, b, Benham et al. 2020), our discovery of 144 regions that may be associated with an adaptive response to environmental variation is not surprising. However, this is the first use of whole genome data to explore local adaptation among subspecies of song sparrows over a microgeographic scale. Our detailed characterization of covariation in the spatial distributions of habitat and genotype at fine geographic scales offers strong support for the hypothesis that selection has contributed to the rapid diversification of locally adapted types in this system. We acknowledge that despite advances in the approaches to identifying GxE associations, there are several challenges with these methods (Hoban et al. 2016). Significant GEAs can alternatively arise from both neutral population genetic and demographic processes (Hoban et al. 2016). Moreover, linked selection via background selection or hitchhiking can result in increased genomic divergence between populations, with loci correlating strongly with environmental variables by chance (Cruickshank and Hahn 2014). To disentangle these processes, our findings warrant future work with a broader spatial and environmental sampling scheme that can help to develop a robust assessment of local adaptation in song sparrows. However, the lack of strong neutral genetic structure among the subspecies studied here, coupled with the absence of IBD based on multiple marker types (Chan and Arcese 2002; Mikles et al. 2020, this study) offers support for our hypothesis of ecological divergence. Moreover, the identification of outliers associated with candidate genes that have previously been linked to tidal marsh adaptions including HSP90B1 (Wan et al. 2017) and PHF20 (Walsh et al. 2019a) provides compelling candidates for future validation. Despite the above caveats, we feel that our work identifies new and important signals of genetic diversity among these populations, which in turn reflects evolutionary distinctiveness of populations that is potentially beneficial to the persistence of these populations both locally and regionally.

Evolutionary distinctiveness and conservation

Our comparisons of putatively adaptive versus neutral loci suggest a pattern of local adaptation and diversity in the song sparrows that highlights the utility of genomic data sets in resolving population-level patterns of diversification. Our findings show that population differentiation is the highest in San Francisco Bay area song sparrows at putatively adaptive loci. Because all five subspecies were well-delineated over three RDA axes, our findings offer empirical evidence of the value of identifying adaptive variation among populations (Crandall et al. 2000; Fraser and Bernatchez 2001). However, despite broad recognition that assessing genetic distinctiveness is a desirable first step in identifying the appropriate units for conservation (e.g., Funk et al. 2012, Volkmann et al. 2014), the appropriate metrics for assessing genome-wide differentiation remains uncertain (Fernandez-Fournier 2021). Although relatively few empirical studies have focused on the conservation of genomic diversity to date, a growing body of literature describes the potentially complementary roles of adaptive and neutral processes in shaping genomic variation within species and its incorporation in conservation planning (Bonin et al. 2007; Funk et al. 2012). Our results generally support these suggestions by elucidating a system in which the inclusion of genomic data on traits putatively under selection revealed aspects of diversity that could be overlooked given nuanced variation in the phenotypes of the subspecies studied here.

More work is needed to determine how patterns of diversity and distinctiveness may affect the adaptive capacity or evolutionary potential of song sparrow populations at micro-spatial to regional scales. However, given operational definitions of adaptive capacity as the ability a species to cope with environmental change (IPCC 2014), and evolutionary potential as an attribute determining a species’ ability to maintain positive long-term growth rates in novel environmental conditions (Thurman et al. 2020), we suggest each will be maximized by conserving units exhibiting novel variation at loci linked to traits underlying additive genetic variance in individual fitness (Hendry et al. 2018). Specifically, our RDA identified several loci putatively linked to salinity and climate, factors driving selection and local adaptation in a variety of taxa (Kingsolver et al. 2012), consistent with the notion that microgeographic variation in selection can increase divergence at local scales whilst reducing it within populations (Hendry et al. 2018; Funk et al. 2019). In the case of song sparrows, isolation and small effective population size (Mikles et al. 2020) might be expected to increase genetic drift among populations adapted to their contemporary environment but compromise their capacity to accommodate change in the future (Funk et al. 2019). However, because our prior results indicate substantial evidence of contemporary gene flow between the five subspecies studied here (Mikles et al. 2020), it is possible that sufficient admixture currently exists among populations to maintain a capacity to respond to variation in natural selection and environment in future. Given an imperfect understanding of these factors at present, we suggest managers prioritize the conservation of evolutionarily significant units (ESUs) defined broadly, based on morphological, genetic, and/or ecological boundaries likely to reflect underlying adaptive process and maintain genetic variation in fitness (reviewed in Funk et al. 2012). Because phenotypic differentiation at fine spatial scales and in response to ecological gradients known to affect individual fitness is a defining trait of song sparrows in North America (e.g., Aldrich 1984, Arcese et al. 2002, Patten & Pruett 2011), we suggest that conserving morphologically and genetically distinct subspecies across the range is a first step towards conserving resilience and persistence in the species overall. Genomic data, interpreted conservatively, can help reveal adaptive and neutral genetic differentiation that can facilitate the prioritization of cryptic species or populations (Coates et al. 2018). When integrated with diagnosable differences in phenotype, as is the case for San Francisco Bay area song sparrows, whole genome data can help delineate populations with high precision.

Revisiting the Song Sparrow Subspecies of the San Francisco Bay

The five subspecies of song sparrow studied here meet the recommended criteria for consideration as distinct populations (McCormack and Maley 2015) given that they were (1) classified a priori by phenotype, and (2) shown to be differentiated at dozens of loci linked to environmental heterogeneity. We now (3) know that they occupy different environmental niches with respect to climate and salinity. Currently, four of these five subspecies (all but gouldii) are listed as ‘species of special concern’ in California, but all song sparrows in California appear to be declining (Sauer 2020). Our use of whole-genome surveys suggests that focusing on adaptive variation can advance management planning in many widespread species with cryptic underlying differences in genetic traits affecting fitness. Quantifying local evolutionary distinctiveness could facilitate predictions on how the influence of climate change, genetic variation, and natural selection may affect potential rates of local evolution (Garant 2020). Moreover, characterizing local adaptive variation may play a role in informing assisted gene flow among these populations, defined as the managed movement of individuals between populations to mitigate local maladaptation (Kelly et al. 2021).