Molecular approach

We collected specimens from much of the range of R. integrifolia including the extremes of the species' distribution plus representative individuals of R. rosea and R. rhodantha. Specimens of R. integrifolia were sequenced from mainland Alaska and St. Lawrence Island in the Bering Sea, south along the Rocky Mountains to New Mexico and along the western coast in the Cascade and Sierra Nevada ranges (Figure 1b). Access to herbaria specimens from the eastern United States was limited, so those populations were undersampled (Figure 1c). Specimens from Siberia were unfortunately unavailable. A total of 35 specimens of R. integrifolia were sampled, representing 27 populations, including those from hypothesized refugial areas, and all known subspecies. Individuals were either collected from the field or from herbarium tissue (see Appendix S1 for Accessions). For the newly collected specimens, collecting permits were provided by the Arctic National Wildlife Refuge (permit #2006-S6) and the Alaska Maritime National Wildlife Refuge (permit 74500-06-015). Freshly collected individuals were preserved in Drierite (WA Hammond Drierite Co. Ltd., Xenia, OH, USA) and voucher specimens were pressed and mounted for curation in the Western Washington Herbarium (WWB; Appendix S1). We isolated genomic DNA from our tissue samples using the DNeasy Plant Extraction kit (Qiagen, Valencia, CA, USA), following the manufacturer’s recommended protocol. DNA was then sequenced from plastid and multiple nuclear loci and subjected to phylogenetic and population genetic analyses to determine the evolutionary history among subspecies and populations.

To place our focal taxa (R. integrifolia subsp. integrifolia, R. integrifolia subsp. leedyi, R. integrifolia subsp. neomexicana, R. rhodantha, and R. rosea) in the broader Rhodiola phylogeny, we performed an analysis using the plastid (pt) spacer region trnL-F, which has been broadly useful for resolving species level phylogenies of many angiosperm clades [26,27]. To accomplish this goal we supplemented our newly sequenced individuals of R. integrifolia, R. rosea, and R. rhodantha with additional GenBank sequences from 17 other species of Rhodiola, two additional sequences of R. rosea, and one outgroup, Cotyledon campanulata Marloth (see Appendix S2 for Accessions). We PCR amplified the pt trnL-F region as follows: each 25 μL reaction volume contained 2.25 mmol/L MgCl2, 0.02 mmol/L dNTPs, 0.05 mmol/L of each primer [26] (Table 1), 2.5 units of Invitrogen (Grand Island, New York, USA) Platinum Taq polymerase, Invitrogen 10X reaction buffer, and approximately 100 ng of genomic DNA and the thermal cycler profile was one cycle at 96°C for 5 min, 35 cycles at 96°C for 45 s, 48°C for 1 min, 72°C for 1 min 30 s, and one extension cycle at 72°C for 10 min. PCR products were sent to the PennState University Nucleic Acid Facility and to Functional Biosciences Inc. (Madison, WI) for enzymatic purification and Sanger sequencing in both directions.

We generated a multi-locus nuclear DNA (nDNA) dataset of anonymous non-coding loci for estimating the evolutionary history within R. integrifolia. The use of multiple nuclear loci in systematic studies greatly improves estimates of species trees and the demographic history of constituent taxa [28]. Non-coding markers were chosen as the basis for our phylogenetic and population genetic analyses because they are i) less likely to be under selection than coding genes, which would confound coalescent-based parameter estimation and ii) more likely to harbor informative nucleotide variation on the shallow timescales relevant to infraspecific studies. Further, our method of identifying loci allowed us to rapidly and randomly describe more non-coding loci and primers than could be obtained from the literature.

Our approach to designing the anonymous locus primer sets (ALPS) for plants was a modification of that described by Galbreath et al. [29] based on Jennings et al. [30], using Invitrogen’s (Carlsbad, CA, USA) TOPO® Shotgun Subcloning Kit. In brief, we i) extracted > 6 micrograms of genomic DNA from leaf tissue of a single representative specimen of R. integrifolia (WWB accession no. 22742) using Qiagen’s DNeasy® Plant Extraction kit, ii) sheared DNA using a nebulizer, iii) blunt-end repaired the fragments, iv) electrophoretically separated DNA fragments on a 0.8% TAE gel, v) extracted fragments ranging from 1 to 1.4 kilobase (kb) in length using Qiagen’s QIAquick® Gel Extraction Kit, vi) added an adenine overhang for ligation into Invitrogen’s pCR®2.1-TOPO® TA vector and insertion into OneShot® TOP10 Competent cells, vii) grew cells on ampicillin-treated plates, and viii) screened colonies using standard blue-white screening. From the plates, we picked and lysed 32 colonies and used the lysate as template in PCR and sequencing reactions using the vector based primer M13. Vector sequence was trimmed from our contigs using Sequencher® v4.8 (Gene Codes Corp, Ann Arbor, MI, USA). To select the appropriate loci for use as ALPS, we ran blastn (nucleotide-nucleotide BLAST) and blastx (translated nucleotide-protein BLAST) searches against the nucleotide databases of the National Center for Biotechnology Information. We discarded DNA fragments matching organelle or protein coding sequences. Furthermore, we removed sequences that had long open reading frames. Fifteen sequences were selected for designing ALPS and primer pairs were optimized in Primer3-web 0.4.0 [31]. These primer sets were tested in temperature gradient PCR reactions to optimize annealing temperatures and to screen for non-target amplification. Ultimately, five ALPS reliably produced clean PCR products across all subspecies of R. integrifolia and were chosen for sequencing (Table 1). It is important to note that because ALPS were designed specifically for R. integrifolia, their utility with other taxa decreased with the degree of evolutionary relatedness. Specifically, the amplification and sequencing of ALPS worked well for all of our R. integrifolia samples, but less so for the distantly related R. rhodantha and R. rosea.

In developing the nDNA dataset for R. integrifolia, our goal was to maximize the number of loci that worked across all populations and subspecies of R. integrifolia rather than the number of individuals because parameter estimation in coalescent analyses improves greatly by increasing the number of independent genetic markers [32]. Therefore, we selected a subsample of individuals that represented R. integrifolia subsp. integrifolia across the regions (Alaska N = 5, Pacific Northwest N=2, Sierra Nevada N=18, Southern Rocky Mountains N=6 [including R. integrifolia subsp. neomexicana and R. integrifolia subsp. procera]), R. integrifolia subsp. leedyi (N = 4), R. rhodantha (N = 2), and R. rosea (N = 7). Nuclear loci were amplified in 20 μl volume reactions, with final reagent concentrations of 1.5 mM MgCl2, 0.5 μM primers, 0.4 mM dNTPs, 0.5 U Taq polymerase, and approximately 5 ng/μl template DNA. All reactions included a 3 min. initial denaturation (94°C), 30 cycles of 30 sec denaturation (94°C), 30 sec annealing (55°C), and 1 min. extension (72°C), and a final 10 min. extension (72°C). All PCR products were cloned using TOPO® TA cloning kits and approximately 10 colonies per individual were used as PCR templates and sent to the PennState University Nucleic Acid Facility and to Functional Biosciences Inc. (Madison, WI) for enzymatic purification and Sanger sequencing in both directions. By taking this approach, we hoped to sample the haplotypes within an individual and thus improve our estimates of the genetic variation within each population cluster (as defined by STRUCTURE, see below).

In preparation for our analyses, all ptDNA and nDNA sequence contigs were assembled and edited using Sequencher® v4.8, aligned using ClustalX 2.0 [33], and alignments were manually edited in MacClade 4.08a [34]. We endeavored to include only orthologs in our analyses, to avoid the "perils of paralogy," such as inferring incorrect gene trees, divergence times, and population sizes [35]. For quality control, we identified paralogous sequences as those with ambiguous alignment or differences in length when compared with other clones within the same individual or across individuals and culled those sequences from our dataset. Furthermore, we manually coded INDELs as simple binary characters following Simmons and Ochoterena [36], and excluded all sites of ambiguous alignment from further analyses.

As a final quality control step for our coalescent-based analyses, we tested each locus for recombination and either removed putative recombinant individuals or reduced the sequence length to the largest region that excluded recombinants. Recombination events confound coalescent analyses that explicitly assume no intralocus recombination by influencing the estimates of demographic parameters [37]. Thus, avoiding recombinants is essential in estimating the historical demography of a species. To test for recombination, we used the RDP [38], MAXCHI and CHIMAERA [39], and GENECOV [40] algorithms implemented in the software package RDP3 [41]. It is generally assumed that recombination is lacking for plastid markers, and is much more likely to encounter a recombinant within nuclear loci. Moreover, because plant nuclear genomes may contain multiple copies of any given gene, the opportunity for recombination in the nuclear genome is much higher [42]. By using the four aforementioned tests of recombination, we were relatively confident in our ability to identify and remove recombinants from downstream analyses.

Using DNASP v. 5 [43], we calculated haplotype diversity (h), the number of segregating sites (θ_S) and nucleotide diversity (π), and performed neutrality tests by comparing Tajima’s D [44] and Ful and Li’s D* and F* [45] statistics to 1000 coalescent simulations of a large, neutrally evolving population.

We then conducted maximum likelihood (ML) analyses in GARLI v0.96 [46] on all loci separately to infer the phylogeny. We performed separate phylogenetic analyses on each of the plastid trnL-F and five nDNA loci using the locus-specific DNA substitution model estimated in jModeltest 0.1.1 [47] (Table 1). In GARLI, we determined the best tree for each locus from ten replicate runs and then assessed branch support based on 200 bootstrap replicates (with five tree searches per replicate). For this, and further analyses we isolated the ptDNA and nDNA datasets in separate analyses because organelle and nDNA may have evolved under different population histories [29,48]. Phylogenies were archived in TreeBASE (http://purl.org/phylo/treebase/phylows/study/TB2:S14743).

The program STRUCTURE 2.3.2 [49] was applied to our multi-locus nDNA dataset to estimate the number of discrete population clusters, rather than relying on pre-defined population structure based on geography or taxonomic descriptions of subspecies. Each haplotype was coded as an allele. For these analyses, we used the admixture model of ancestry and both the correlated [50] and uncorrelated allele frequencies models, running separate analyses for all values of K (number of population clusters) between 1 and 7 (accounting for the overall number of disjunct regions and subspecies sampled). After a burn-in of 100,000 generations, we ran the analysis for an additional 100,000 generations and repeated them twice to confirm reproducibility.

The phylogenetic relationships among the four major population clusters of R. integrifolia (assigning individuals to regional groups based on the output from the STRUCTURE analyses), and R. rhodantha, using R. rosea as an outgroup (as determined by our ptDNA analyses of relationships within Rhodiola) were inferred in the program BEAST 1.6.2 [51], as implemented using *BEAST [52]. In doing so, all five of the nuclear loci were used in a single analysis. This coalescent-based approach produces more reliable multi-locus estimates of species trees and divergence times because alleles are not arbitrarily linked among loci as they might be in concatenated analyses [53]. We applied the Yule tree prior, given that we were estimating the relationships among several species and subspecies [51], and allowed rates to vary among loci. The ptDNA and nDNA analyses were run for 100 and 500 million generations, respectively, with 10% of each run discarded as burn-in. Stationarity was determined by examining parameter trend plots and effective sample size (ESS) values (all > 200) using TRACER 1.5 [54]. All analyses were repeated three times using different random seeds to confirm convergence of parameter estimates. To calculate divergence times, we used two different estimates of mutation rate: i) μ based on trnL-F 8.24e-9 substitutions/site/year calibrated from fossils of Aichryson Webb & Berthel. (Crassulaceae) [55] and ii) a general μ for plant autosomal genes of 1.1e-8 substitutions/site/year [56,57]. The resulting dates should be viewed with caution, given that i) Aichryson is distantly related to Rhodiola and ii) the autosomal mutation rate may be different from that of non-coding loci.

The history of population size and gene flow was quantified simultaneously based on the nDNA dataset using the Bayesian formulation of LAMARC 2.1.6 [58]. For the demographic analyses, we applied the locus-specific models of nucleotide evolution from jModeltest [47], the estimates of relative mutation rates among loci from *BEAST, retained the default priors for θ (shape = logarithmic, upper bound = 10, lower bound = 1e-05) and migration (M), and adjusted the prior on the growth parameter (g) to a uniform distribution spanning -1000 to 2000. Final analyses were run for 100 million generations following 20 million generations of burn-in, sampling every 5000 generations. We assessed parameter trend plots, ESS values, and the shape of posterior distributions using TRACER to confirm stationarity and three separate runs were conducted to confirm repeatability of our results.

Ecological niche models

Based on our molecular analyses, we evaluated differences in the distribution occupied by the two major R. integrifolia clades identified in our analyses–i.e., Beringia and the Southern Rocky Mountain Refugium (SRMR; excluding R. rhodantha; see results below)-by generating ecological niche models. First, the potential distribution of each clade was estimated based solely on occurrence records within that clade and then those were compared to visualize discrepancies in the capacity of records from one clade to predict the overall distribution of the lineage. Then, principle components analysis (PCA) and tests of niche equivalency and similarity were used to assess differences in the environmental space occupied by each clade. Our ecological niche models were built using georeferenced herbaria records (Figure 1a-c, Appendix S3) and CRU TS 2.1 climate data for the period 1971-2000 at its native 0.5° resolution [59]. The Beringia clade consisted of all records for all subspecies of R. integrifolia except those located in Colorado and New Mexico (n = 272 with one record used per pixel). The SRMR clade was composed of all records for Colorado and New Mexico (n = 39).

Predictor variables were selected using the random forest (RF) algorithm [60] based on recommendations by Strobl and colleagues [61]. Six variables were selected from an initial group of 44, based on RF output and Spearman’s correlation values (variables correlated at rho values ≤ |0.8|): annual temperature range (maximum temperature of the warmest month minus minimum temperature of the coldest month); mean diurnal range for summer and winter (mean of the maximum temperature minus minimum temperature of each month for June–August and December–February); summer minimum temperature (mean of minimum temperatures for June–August); precipitation of the wettest month; and cumulative spring snowpack (sum of snowpack for March–May). These variables correspond to climatic factors that are especially relevant to arctic-alpine plants [62].

An ensemble modeling approach was used to build ecological niche models. Ensemble modeling is based on the idea that different combinations of initial conditions, model algorithms, and general circulation models (GCMs) represent alternate possible states of the system being modeled [63]. When these are combined using consensus methods (such as the mean of all models), they can form a more accurate projection, outperforming single models [64]. For temporal projection, the two greatest sources of uncertainty in SDM models are the choice of model algorithm and the GCM used [65]. For this reason, the use of single algorithms and GCMs to project changes in species distributions has been criticized, especially for paleodistribution modeling [66]. In this study, we use an ensemble of multiple model algorithms and GCMs.

Modeling was conducted with the BIOMOD package [67] in R v. 2.13.1 [68] and Maxent, v. 3.3.3 [69]. Eight model algorithms were used: generalized linear models, generalized additive models, multivariate adaptive regression splines, classification tree analysis, flexible discriminant analysis, generalized boosted models, random forest, and maximum entropy (Maxent). All models except Maxent were run using R. For all models the entire background of the study area was used to create pseudo-absence data, with these data weighted to maintain the number of presence records equal to the number of background records (prevalence = 0.5).

Model performance was assessed using ten bootstrapped replicates for each algorithm, partitioning the data into 70% for training and 30% for verification. The area under the curve of the receiver operating characteristic (AUC [70]) was averaged across replicates and used to determine if models should be removed from the ensemble. Models with mean AUC > 0.7 were determined to be useful and were kept in the consensus analysis. Final models were built using 100% of the available data [71]. Model probabilities were converted to presence/absence predictions using two thresholds: minimizing the absolute value of sensitivity minus specificity, and the mean probability value across model output. Both of these methods have been shown to perform well in threshold comparisons [72]. Current distributions represented the percentage of models voting “suitable habitat” for a given pixel based on 16 model votes (eight algorithms x two threshold methods). Final maps for each clade/species were based on 75% or more of the models voting suitable habitat for a given pixel.

Paleodistribution models were built using the combined R. integrifolia records (all individuals from both the Beringia and SRMR clades). Models were calibrated and evaluated as above and then projected onto climate data from the Mid-Holocene (6,000 years before present), the Last Glacial Maximum (LGM, 21,000 years before present), and the Last Interglacial (LIG, 124,000 years before present), which were assumed to serve as proxies for glacial and interglacial stages throughout the Pleistocene. Paleoclimate data were interpolated from their native resolution to 0.5° resolution using ordinary cokriging in ArcMap [73] using the CRU TS 2.1 climate data as the secondary data set. Mid-Holocene and LGM climate data were derived from three GCMs available through the Paleoclimate Modeling Intercomparison Project Phase 2 [74]: CCSM 3 [75], HadCM3 [76], and MIROC 3.2 [77]. Data for the LIG were provided through the HadCM3 model. Paleodistributions for the Mid-Holocene and LGM represent consensus across these three GCMs. The final refugial map was derived by assessing the probability of suitable habitat in each pixel across all four time slices (including the current distribution). The refugial map indicates pixels classified as suitable habitat across these time slices based on two thresholds for suitable habitat, 50 and 75%. The LGM ice layer [78] was overlaid to exclude potential suitable habitat under the ice sheet and create the final map of potential refugia.

Principle components analysis (PCA) was used to evaluate the niche space occupied by the two focal clades of R. integrifolia based on climate variables (as described above) located by georeferenced herbarium samples. PCA was conducted in the context of the available environmental space, as recommended by Broennimann et al. [79]. Climate data were centered and standardized prior to PCA analysis; analyses were conducted in R using the ade4 package [80]. Overlap between the two clades in niche space was calculated using the D metric [79], which ranges between 0 (no overlap) and 1 (full overlap); randomization tests were used to assess niche equivalency and similarity. The null hypothesis for niche equivalency is that the clades have equivalent niches, while the null hypothesis for niche similarity is that the clades have retained niche similarity. Visualization of PCA analysis and calculation of D, equivalency and similarity were conducted with R code provided by O. Broennimann (Univ. Lausanne; further description of these tests can be found in [79]).

Relationships within Rhodiola

Our analyses have untangled the evolutionary history of R. integrifolia, helping shed light on taxa that are disjunctly distributed across the arctic-alpine biome. Our trnL-F phylogeny indicates that the North American R. integrifolia and R. rhodantha are more closely related to one another than either is to R. rosea. Indeed, R. rosea was placed in a separate clade of Asian taxa. Furthermore, this phylogeny indicated that some populations of R. integrifolia are more closely related to R. rhodantha than to other populations of R. integrifolia. The more rapidly evolving ALPS broadly confirmed our findings from these plastid analyses, in addition to clarifying the demographic history within R. integrifolia. Inferences from the five nDNA loci revealed two major clades, based in Beringia and the SRMR, that diverged ~ 700 kya (660 kya or 880 kya depending on the mutation rate [Figure 5]) and have remained isolated from each another. The close ties between the SRMR clade and R. rhodantha, with a divergence ~ 500 kya (or 660 kya depending on the mutation rate [Figure 5]), lends support to Clausen’s [17] and Guest's [15] hypotheses that R. integrifolia and R. rhodantha are closely related. The novelty from our study is that R. integrifolia has experienced a complex evolutionary history of isolation among multiple disjunct refugia since the mid-Pleistocene that gave rise to several lineages endemic to North America.

The association between R. integrifolia and R. rhodantha evident in the tree topology could have arisen i) as the result of hybridization and introgression where the two species of Rhodiola overlap or ii) because R. rhodantha and the SRMR R. integrifolia clade are closely related. Though the chromosome counts and several characteristics (e.g., dioecy/hermaphrodite, flower color) differ between R. rhodantha and these southern R. integrifolia, they do share the common morphological feature of being taller plants in general, distinguishing the SRMR populations from R. integrifolia to the north. The hybridization scenario (i) likely involved an ancient split between R. integrifolia and R. rhodantha from a common ancestor as envisaged by Clausen [17]. Given the chromosomal and morphological differences between the species, the common occurrence of hybridization within the Crassulaceae [82], and the geographically sympatric distributions, the ancient split from a common ancestor [17] followed by more recent hybridization in the south between the two distinct species is a plausible scenario [15]. These analyses reveal that the evolutionary histories of R. integrifolia and R. rhodantha are tightly linked, but a more thorough analysis of members within this SRMR clade, including additional chromosome counts, are needed in order to more fully elucidate their history.

Two geographically disjunct clades within R. integrifolia

We identified two major clades of R. integrifolia in this study, Beringia and SRMR, that diverged ~ 700 kya (660 kya or 880 kya depending on the mutation rate [Figure 5]). This immediately followed the Middle Pleistocene Transition (MPT) [83], and since then there has been extremely limited gene flow between the two clades. The MPT marked the transition towards an increase in the severity of glaciations, the emergence of the ~100-kyr glacial cycles (MPT beginning ~ 1250 kya [84]), and maximum glaciations and warmer interglacials becoming established at the end of the MPT ~700 kya [85]. Sedum lanceolatum also exhibited a dramatic phylogeographic break between the central and southern Rockies, across the low elevation and xeric habitat of the Wyoming Basin [86], during the same time-period (~600 kya [3,87]), even though it is not restricted to tundra habitats. Together, evidence from these species within the Crassulaceae suggests that the MPT had a profound impact on the distribution and diversification of arctic and alpine plants.

Isolation in these two widely disjunct refugia is also independently supported by paleodistribution models, which indicate major refugial areas in Beringia, the Sierra Nevada Mountains and the southern Oregon Cascades, and the SRMR since the Last Interglacial (Figure 7d). Niche models based on current distributions of the Beringia and SRMR clades show that these models have a moderate ability to predict one another (Figure 1d) indicating that the clades continue to share some parts of their climatic niches, despite apparent genetic isolation between these lineages. This is also supported by PCA analysis (Figure 6), which shows very little overlap on PC axis 1 (defined by minimum summer temperature and summer mean diurnal range) and complete overlap on PC axis 2 (defined by spring snowpack and precipitation of the wettest month). Thus, based on the predictors we used in niche modeling, we see separation between the two clades based on minimum summer temperature and summer mean diurnal range. While statistical tests of the niches of the two clades indicate that they are not equivalent, the niches are more similar to each other than random expectation, supporting a likely role of geographic isolation in promoting genetic divergence in these taxa.

Conservation implications

An improved understanding of systematics and the roles of refugia in maintaining and promoting diversity are central to understanding the process of speciation and informing conservation priorities [81]. The evolutionary history revealed by our multi-locus phylogeographic analyses shows that multiple genetically distinct and narrowly endemic R. integrifolia populations exist in North America, some of which are of conservation concern. The environmental differences among the disjunct refugia and the indication of diversification of ecological niches in these clades highlights the need to preserve these regions that harbor genetic variation and support processes of genetic divergence. Thus, these areas are important target areas for conservation under global change [115], and the demographic history of R. integrifolia should serve as a rationale for similar studies in other plant species.

Alpine species, with distributions defined by steep climatic gradients, are being disproportionately impacted by ongoing climate change [116,117,118,119,120]. In alpine and arctic systems, range shifts in response to climate change will likely result in increasingly fragmented and reduced habitat as species move higher in elevation and latitude, assuming that migration corridors remain intact. The consequence for many local populations will be extirpation as they run out of high elevation and high latitude habitat. Given the highly reduced and fragmented habitat available in the east and the scarcity of populations of R. integrifolia subsp. leedyi, this lineage is by far the most at risk from the changing environmental conditions associated with warming. Including refugial locations, especially those in the south or 'rear edge' [121], in conservation planning can help build evolutionary resiliency within populations, mediating extinction risks [122,123].