2.1 Study area and focal species

The present study is focused on the NE Pacific coast, along the distributional ranges of P. torreyi and P. scouleri, from the Baja California Peninsula to Alaska. Samples of P. torreyi were obtained from San Juanico (26.2 N, 112.4 W, Mexico) to Fogarty Creek (44.8 N, 124.1 W, Oregon) and P. scouleri was sampled from Government Point (34.4 N, 120.5 W, California) to Vancouver Island (48.6 N, 124.5 W, Canada) (Table S1). To further check possible admixture between east and west Pacific Phyllospadix populations, we sampled three populations of P. iwatensis in Hokkaido Island (Japan; Table S2). Across sampling areas, approximately 30 individuals were sampled haphazardly at low tide along approximately 30 m, keeping a minimum distance of approximately 1 m between each sampling unit at each location along 23 sites (see Tables S1 and S2). All sampled plants were blotted dry and preserved dry in silica gel.

2.2 Species distributions models

Species Distribution Modelling (SDMs) were built for P. torreyi and P. scouleri with two machine learning algorithms, Boosted Regression Trees (BRT) and adaptive boosting (AdaBoost). These are well-suited for modelling complex interactions and non-linear relationships and have been shown to have high predictive performance while reducing overfitting through proper hyper-parametrization and forcing of monotonic responses (Elith & Leathwick, 2007; Hofner et al., 2011).

Biologically meaningful predictor variables for seagrasses (Krause-Jensen et al., 2020), namely temperature, nitrate, salinity and sea-ice, were extracted from Bio-ORACLE (Assis, Tyberghein, et al., 2018), a dataset providing global marine data layers for ecological modelling at a 0.08° resolution (approx. 9 km at the equator). These predictor variables were obtained for present-day conditions, the Last Glacial Maximum (LGM), and the extreme Representative Concentration Pathway (RCP 8.5) future scenario of increased carbon emissions with no climate mitigation policies, as detailed in a previous publication utilizing the Bio-ORACLE pipelines (Assis et al., 2023) (see Data availability statement). The choice of RCP 8.5 helps to estimate the potential impact of the most extreme predicted climate change scenario on species' distributions (Assis et al., 2022; Gouvêa et al., 2022). The predictors chosen reflected essential resources (nutrients as nitrate) as well as factors affecting physiology (salinity and temperature) and perturbation (sea ice cover). The LGM predictors considered the weighting of different Atmospheric-Ocean General Circulation Models (AOGCM; Assis, Tyberghein, et al., 2018) as well as the potential −120 m sea-level change during the LGM (Assis et al., 2016, 2018; Peltier & Solheim, 2004).

Because the focal species are intertidal, predictors were clipped to the regions defined by NE Pacific coastlines. Presence records for the two species were extracted from two biodiversity information facilities, GBIF (2021a, 2021b) and OBIS (https://www.obis.org/), and from the fine-tuned dataset of marine forest species (Assis et al., 2020). An equal number of pseudo-absences was randomly generated in locations where no presences were recorded (Assis, Serrão, et al., 2018). To reduce the potential effect of autocorrelation and sampling bias in the models, we filtered the records based on the minimum distance at which predictors were not significantly correlated (Figures S7–S9).

This distance was inferred by building a correlogram per species that identified the variability of Pearson's correlation of predictors with increasing marine distances (Krause-Jensen et al., 2020). This process reduced the initial 941 records of P. scouleri to 195, and the initial 568 records of P. torreyi to 87.

A 6-fold cross-validation approach was implemented, where independent blocks were used to find the optimal combination of hyperparameters, specifically, the number of trees (ranging from 50 to 1000, step 50), tree complexity (ranging from 1 to 6), and learning rate (0.01, 0.005, and 0.001) for BRT (Elith & Leathwick, 2007) and shrinkage (from 0.25 to 1, step 0.25), degrees of freedom (from 1 to 12, step 1) and number of interactions (from 50 to 250, step 50) for AdaBoost (Hofner et al., 2011). The performance of the model was evaluated with the area under the curve (AUC) of the receiver operating characteristic curve and True Skill Statistics (TSS), which is determined by the sum of sensitivity (true presence rate) and specificity (true absence rate) minus one (Allouche et al., 2006) and the Boyce index, which is a proper metric for presence-pseudo-absence models (Boyce et al., 2002). To reduce overfitting, negative monotonic responses were forced for maximum temperature and ice thickness, and positive responses for nitrate, salinity and minimum temperature (Hofner et al., 2011). By implementing such modelling constraints, the main hypotheses are (1) for predictor variables forced with negative responses (e.g., maximum temperature), there is an extreme maximum value from which species' probability of occurrence is zero, and (2) for predictor variables forced with positive responses (e.g., minimum temperature), there is an extreme minimum value below which species' probability of occurrence is zero. The relative importance of each predictor was assessed by measuring the increase in AUC when adding each predictor to the alternative model (Assis, Serrão, et al., 2018).

To develop parsimonious models per algorithm, with higher transferability potential, a stepwise approach was used, starting with a full model, and iteratively removing predictors from the least to the most contributory until the reduction in AUC was significant. This step was performed using a Wilcoxon Rank Sum test on the AUC values generated in the cross-validation procedure. Maps reflecting species' potential distribution for the present-day, the LGM and future conditions were created by averaging (i.e., ensemble modelling; Araújo & New, 2007) the responses of both BRT and AdaBoost parsimonious models. These were then reclassified into binomial maps showing the presence or absence of both species by using a threshold that maximized TSS (Allouche et al., 2006). To assess for past range shifts, we compared the modelled distributions for the LGM and the present and identified refugial areas where the species was predicted to occur in both time periods.

2.3 Primer design, microsatellite amplification and genotyping

Genomic DNA for microsatellite development was extracted from 500 mg of dried tissue from a pool of 97 P. torreyi individuals and 500 mg from 66 P. scouleri individual samples. These samples were used to run 454 sequencing (Biocant, Cantanhede, Portugal). Simple di- to hexanucleotide microsatellite loci motifs were tested for primer design using Primer 3 (http://biotools.umassmed.edu/bioapps/primer3_www.cgi). A total set of 38 and 27 pairs of primers were designed for P. torreyi and P. scouleri, respectively, and tested for amplification. To determine the applicability and amplification success of the microsatellites, we tested them on two individual samples from five populations each of P. torreyi and P. scouleri. From the initial set, a subset of 38 loci for P. torreyi and 20 for P. scouleri, respectively, were chosen due to their amplification success. These were evaluated for their ease of scoring and polymorphism using primers labelled with fluorophores, and 11 were finally selected due to their number of repeats, similar melting temperature allowing multiplexing and cross-amplification success (primer sequences and microsatellite motifs listed in Table S3, supplementary material).

Genomic DNA was extracted using NucleoSpin96 PlantKit II (Macherey-Nagel, Germany). PCR reactions were performed in a total volume of 15 μL, containing 1× Colourless GoTaq Flexi Buffer (Promega, Madison, WI, USA), 2 mM MgCl₂, 10 μM forward and reverse primers, 0.2 mM of dNTPs, 1U GoTaq G2 Flexi DNA Polymerase (Promega, USA) and 5 μL of template DNA. Cycling conditions consisted of an initial denaturation step at 95°C for 2 min, followed by 30 cycles of 95°C for 30 s, annealing temperature (T _a, Table S3) for 30 s, extension at 72°C for 30 s and a final extension at 72°C for 5 min. Amplified fragments were combined with 0.25 μL GeneScanTM 500 LIZ Size Standard (Applied Biosystems, UK) and 9.75 μL of Hi-Di formamide, followed by 5 min denaturation at 95°C and their analysis were performed on an ABI3130xl automated DNA sequencer (Applied Biosystems, Waltham, MA, USA).

Although P. torreyi and P. scouleri were genotyped for 11 microsatellite loci (Table S3), P. iwatensis populations were only genotyped for a set of eight loci (Psc07, Pto07, Pto13, Pto15, Pto25, Pto29, Pto31 and Pto35) as amplification failed for the additional three loci.

2.4 Microsatellite data analysis

Fragments were scored using STRAND (Veterinary Genetics Laboratory, University of California, Davis; http://www.vgl.ucdavis.edu/STRand) and binned using the R package ‘MsatAllele’ (Alberto, 2009). Prior to analysis, the final dataset was run in Micro-Checker 2.2.3 software (van Oosterhout et al., 2004) to test for the presence of null alleles and test Hardy–Weinberg equilibrium and run a linkage disequilibrium test. No evidence of null alleles or linkage disequilibrium was found for any of the 11 loci; therefore, they were all considered suitable markers for the analyses of population structure.

Seagrass species occupy space via rhizome growth, a type of clonal growth where genets (i.e., genetic individuals coming from distinct seeds) form ramets (i.e., units capable of independent life if separated, formed by shoots, associated roots, and a piece of rhizome). Therefore, when seagrasses are sampled, duplicate genotypes (clones) may be collected that derive from a single seed. To assess the clonality of each population, the package RClone (Bailleul et al., 2016) was used to assess the probability that identical multi-locus genotypes (MLGs) are the product of the same or separate reproductive events. The probability of finding identical MLGs resulting from distinct sexual reproductive events (Psex) was computed for each population (Arnaud-Haond et al., 2007) to check if individuals with the same MLG were truly clones. When Psex was lower than the threshold value (fixed at 0.01), two identical MLGs were considered to be the same clone. Based on the number of individuals (N) and genets (G) sampled, clonal diversity in each population was calculated using the clonal richness index (R), (R = [G − 1]/[N − 1]), ranging from 0 (one single clone) to 1 (when all samples analysed correspond to a different genet). Only unique genets per population were used for genetic diversity analyses to avoid biasing allelic frequencies, that for the research questions of this study, should represent the population of genets rather than the population of ramets.

2.5 Genetic diversity and population genetic structure

To compare population genetic diversity and differentiation, all populations must have the same loci amplified. Two datasets were therefore constructed for genetic analysis as only eight loci amplified for P. iwatensis out of the 11 that were used to amplify P. torreyi and P. scouleri. Dataset 1 includes P. torreyi and P. scouleri populations (11 loci), and dataset 2 encompasses populations for all three species (eight loci). Additionally, for higher resolution analysis of population differentiation within each species, separate genetic analyses were performed for species in dataset 1 individually (P. torreyi and P. scouleri). A total of 593 Phyllospadix samples were genotyped (304 P. torreyi and 289 P. scouleri samples) for dataset 1 (11 loci).

Summary statistics of genetic diversity were calculated for each population using GENETIX 4.05 (Laboratoire Génome, Populations, Interactions, Université de Montpellier II; http://kimura.univ-montp2.fr/genetix). This included the mean number of alleles per locus (allelic richness), the non-biased expected heterozygosity (H _E), the observed heterozygosity (H _O), the number of private alleles, and the multi-locus inbreeding coefficients (F _IS). Sample sizes were also standardized to the smallest in any population using 10⁴ randomizations.

Population genetic structure was inferred using STRUCTURE 2.3 (Pritchard Lab, Stanford University; http://pritchardlab.stanford.edu/structure.html) without any prior population assignments. Structure was run for both datasets and for P. torreyi and P. scouleri individually. The estimated number of genetic clusters, K, was estimated from the mean log likelihood for each K value and the ΔK statistic (Earl, 2012) to identify the optimal number of groups. This was analysed considering a range of K values from 1 to 23. Ten replicates per K value were used to verify the consistency with a 50,000 burn-in followed by 500,000 MCMC replicates per iteration. All jobs were run in parallel on multiple cores using the R package ‘ParallelStructure’ (Besnier & Glover, 2013). The best K was estimated using the program Structure Harvester which ranks K's according to Evanno method (Earl, 2012). Structure analyses were complemented with a factorial correspondence analysis (FCA) implemented in GENETIX 4.05.2 (Belkhir et al., 1996). To quantify pairwise differentiation between populations, we calculated F _ST and Jost's D distances using the software GENODIVE (Meirmans & van Tienderen, 2004).

3.1 Species distribution models

The ensemble of both BRT and AdBoost algorithms resulted in high performance predictions, largely matching the known distribution of the two species (Figure 1; maps depicting the probability of occurrence available in Figures S10 and S11). Specifically, for P. scouleri AUC was 0.99, TSS was 0.93 and Boyce was 0.89, while for P. torreyi AUC was 0.99, TSS was 0.91 and Boyce was 0.73. The models considered numerous biologically meaningful predictor variables, from which temperature (maximum and minimum) were key in explaining the distribution of both species (Figures S1, S2, S10 and S11), in line with additional studies based on species distribution modelling of seagrasses (Gouvêa et al., 2024).

The models hindcasted significant equatorward range shifts during the Last Glacial Maximum (LGM). The extreme colder environments might have contracted poleward range margins down to Oregon (P. scouleri) and California on the West coast of the USA (P. torreyi). Concomitantly, lower latitude range margins might have remained stable or even expanded along Central American latitudes (Figure 1). In contrast, the future scenario of increased carbon emissions (RCP 8.5) projected poleward range expansions with potential habitat losses at the lower latitude range margins of Baja California Sur for both species (Figure 1).

3.2 Population genetic and genotypic diversity

Global and endemic genetic diversity generally decreased from north to south for P. torreyi but not for P. scouleri (Figure 1; Table S1). Local diversity (private alleles) ranged from high in southern populations of P. torreyi to an almost complete lack of endemism in the most northern populations (Figure 1a). The two species differed particularly in their contrasting northernmost range, where P. torreyi displayed very low within-population diversities whereas P. scouleri showed some of the highest diversity metrics in the north (Table S1).

More than half of the populations from both species had significant heterozygote excess relative to the expectation under random mating (Table S1, significant negative F _IS). For dataset 2, P. iwatensis revealed lower genetic diversity relative to the two eastern species (e.g., mean A = 2.5; Â ₇ = 2.06 ± 0.1 and PÂ = 0.37 ± 0.75; Table S2) and significant heterozygote excess (negative F _IS) in all populations (Table S2).

When P _sex values were higher than 0.01, the identical MLGs were considered derived from independent sexual reproduction events (i.e., distinct seeds), resulting in a total of 494 unique genotypes for dataset 1 (G = 255 for P. torreyi and G = 239 for P. scouleri; Table S1). The mean global genotypic richness was relatively high for both species (R = 0.88 and R = 0.92; Table S1), indicating predominant sexual reproduction. The western species P. iwatensis had slightly lower clonal richness, but still a larger proportion of the sampled shoots were derived from sexual reproduction rather than clonal propagation (R = 0.74, Table S2).

3.3 Genetic differentiation and structure

The structure results for the three species combined (dataset 2) showed differentiation between the three species, the same genetic structure for P. torreyi and P. scouleri (as in dataset 1) and P. iwatensis forming a distinctive cluster. The northernmost population of the eastern Pacific (Vancouver, P. scouleri) had individuals with a proportion of the genome assigned to the same group as the western Pacific species, P. iwatensis, in addition to the admixture already previously identified with its more southerly sister species, P. torreyi (Figure 2). The FCA retrieved similar results supporting the differentiation between the western versus eastern species along the first axis, and among the two eastern species along the second axis (Figure 2).

Structure analysis including all Eastern Pacific samples of both species revealed, at the first hierarchal level, a separation between two species (Figure S3). However, some populations morphologically identified as P. scouleri showed genetic admixture from P. torreyi (Figures S3 and S4), specifically at the two range extremes of P. scouleri (its southernmost and northernmost limits) and at an additional site (San Simeon) in the middle of the distribution of both species.

The subsequent hierarchical level of genetic structure (K = 5) divided south, centre, and northern populations for both species (Figure 1a,b,d,e), with some admixture as previously shown at the main split into two species K = 2. The same three groups per species were corroborated when structure analyses were run for each species separately (Figure 1a,e). P. torreyi populations were divided into (1) Mexico group (south cluster); (2) South California group (centre cluster); and (3) North California/Oregon group (north cluster) (Figure 1a,b). The same number of clusters were found for P. scouleri: (1) South California (south cluster); (2) Centre California (centre cluster); and (3) Oregon/Vancouver (north cluster), with south and centre clusters having an intermingled transition region (Figure 1d,e).

All populations within each species were significantly differentiated (p < 0.01) for both F _ST and Jost'D estimates (Figures S5 and S6). The pairwise differentiation values within P. torreyi and P. scouleri were lower between southern populations, whereas the northern populations against all others had the highest differentiation (Figures S5 and S6).

4.1 Genetic signatures left by past climatic changes

This study shows that the expectations of effects of past range shifts inferred by distribution models can be found in the current genetic variability across the distributional range of surfgrass species P. torreyi and shows that the same patterns were not observed in the sister species P. scouleri that has a colder affinity range distribution (as illustrated in the models of ranges). The contrasting results observed between species are a novel interesting result of this study, as the small thermal affinity differences between them correspond to large differences in private genetic diversity, with many unique alleles having accumulated in the southern region. Populations able to persist in any of these refugial areas should have accumulated a larger amount of new private genetic diversity and genetic distances across populations, in contrast with regions that suffered local extinctions followed by recent settlement (Jacobs et al., 2004; Mann & Hamilton, 1995). This is why private alleles are the important variable to locate regions of long-term persistence to reflect past stability rather than contemporary influences.

In contrast to northern populations, the southern edge populations of P. torreyi, located on the Baja California Peninsula in Mexico, were more genetically unique (private alleles, Table S1) and differentiated (FCA, Figure 1). The decline in the private alleles with increasing latitude contrasts with expected heterozygosity, particularly in the southernmost Baja California population which is less diverse (Table S1). Higher diversity in more ancient populations at southern range margins is expected in temperate species of the northern hemisphere, as long-term population persistence may have been critical in preserving and accumulating genetic differences (Hampe & Petit, 2005; Hewitt, 1999, 2004) with no notable effects from drift, local extinctions or bottlenecks. Population genetic diversity decreasing from lower to higher latitudes corroborates patterns in other intertidal vegetation species (e.g., Assis, Serrão, et al., 2018; Neiva et al., 2015) and across subtidal marine vegetation as well (e.g., Assis et al., 2016; Diekmann & Serrao, 2012; Johansson et al., 2015; Neiva et al., 2020), all likely to have been affected by the glacial cycles at higher latitudes.

Despite the expectation of lower genetic diversity at higher latitudes (Eckert et al., 2008; Excoffier et al., 2009) due to genetic bottlenecks associated with expansion after colder periods (Marko, 2004), this pattern was not observed in P. scouleri. We proposed and analysed two hypotheses that could explain the high diversity at its northernmost range: (1) admixture of genetically divergent entities such as introgression with western Pacific species of Phyllospadix, or (2) long-term persistence in this area, which might have remained ice-free up until the late Pleistocene when the glaciers retreated (Jacobs et al., 2004; Mann & Hamilton, 1995). The northernmost population analysed herein revealed higher private allelic richness but also shared genome assignment with the western Pacific group, suggesting admixture, thereby supporting the hypotheses of hybridization. However, it's important to note that the evidence for hybridization is somewhat weak, with only a small proportion of individuals showing evidence of hybridization with P. iwatensis. On the other hand, the distribution models suggest the possibility of persistence somewhere nearby, as the expected distribution of P. scouleri during the LGM closely aligns with the location of the northernmost population. Therefore, while hybridization remains a possibility, the evidence also hints at the possibility of long-term persistence nearby. The alternative hypothesis of incomplete lineage sorting (ILS) is less supported by the evidence because ILS is not expected to affect only the northernmost edge population. This pattern has been found for other NE Pacific species (Carrara et al., 2007; Jacobs et al., 2004).

4.2 Future predictions

The Intergovernmental Panel on Climate Change (IPCC) reported that the global temperature is expected to surpass 1.5°C above pre-industrial levels in the next decades, potentially exposing population and species to detrimental conditions (IPCC, 2022). Accordingly, local population extinctions, shifts, or losses of unique genetic diversity are projected (Chefaoui et al., 2018; Gouvêa et al., 2024). Specifically for our model species, recent observations show that the warming trend of the Baja California peninsula, Mexico, is in line with the RCP 8.5 scenario (Martínez-Austria & Jano-Pérez, 2021). Our projections estimate that both Phyllospadix species might have similar trailing edges in the near future as a consequence of range shifts. Due to unsuitable conditions throughout the southern part of the Baja California peninsula, the models project a northward shift of the trailing edge to the Northern more temperate region of Baja California (~30° N) by the end of the century. Baja California Sur, showing high present-day endemic (i.e., private alleles) genetic diversity, lining up with the hypothesis of providing a climatic refugium, is predicted to eventually lose populations there, resulting in a climatic impact of the loss of the global genetic diversity pool of P. torreyi. Such loss, dependent on the pace of future range changes, would be irreversible. Moreover, private alleles generally support the potential for species to undergo selection and thereby adapt to future environmental changes, as well as to avoid inbreeding depression (Hampe & Petit, 2005; Petit et al., 1998). Additionally, considering the role of Phyllospadix as a foundation species, the regional loss of populations may generate important negative cascading effects on the entire associated tidal community (Shelton, 2010). Therefore, the southern populations of P. torreyi have important priority conservation value for hosting the most unique genetic diversity found in this study.

On the other hand, the projected northward shifts in the suitable habitats of both Phyllospadix species may allow the colonization of new areas, particularly along Canada and Alaska. The projections for the leading-edge differed between both species, with P. torreyi potentially experiencing greater poleward expansions along the Gulf of Alaska (to ~57° N) and spreading to new colder and higher-latitude areas with unknown but predictable possible founder effects creating novel mosaics of genetic diversity and structure. In contrast, for P. scouleri, the poleward expansion is projected to be less extensive along this coastline, but its colder tolerance suggests that it might hypothetically find opportunities to expand further west into Russian coastlines, although such habitats are not part of the focus of the present study.

Based on projections, neither Phyllospadix species is in danger of extinction, despite the previously mentioned negative effects on their genetic diversity levels and localized loss of ecosystem services. However, because only a part of the genetic diversity is expected to shift, the spread into new regions may result in a high frequency of particular genetic variants due to allele surfing (Excoffier & Ray, 2008; Neiva et al., 2010). Species may be able to migrate and follow suitable conditions that better fit their niche if the pace of climate change allows (Ackerly, 2003; Hewitt, 2000). This process will result in a gradient of increasing genetic diversity toward the leading edge (McInerny et al., 2009). Moreover, it is important to point out that given the short time frame under the RCP 8.5 projection (i.e., until 2100), species that might have limited dispersal traits, such as the Phyllospadix genus (Kendrick et al., 2012), will require a larger amount of time for effective gene flow, increasing the likelihood that alleles become extinct before reaching new areas (Knutsen et al., 2013).