The sampling sites covered a gradient from the Alps to Sicily and to Greece (Table 1 and Figure 1). Populations on the Apennines (ITC1, ITC2, ITC3, ITC4, ITC5, ITC6, ITS1, and ITS2) and southern Rhodopes in Greece (GR1 and GR2) are MaP populations disjunct from the main continuous distribution range. Populations on the Alps (ITN1, ITN2, and ITN3) and Balkans (SRB) represent populations in the southern main distribution area. In order to assess if geographical peripherality was related to environmental limitations, an ensemble of current climatic data was used to analyze the environmental variability of the study sites. The ClimateEU software¹ was used. This software extracts and downscales Parameter-elevation Regressions on Independent Slopes Model (PRISM) monthly data (0.5° resolution) to scale-free point data (Daly et al., 2008) and calculates seasonal and annual climate variables for specific locations based on latitude, longitude, and elevation. The software was used to generate 15 biologically relevant climatic variables for the 1981–2010 normal period (Supplementary Table S1) and with 250 m of spatial resolution. A standardization procedure was applied as the variables are expressed in different scales. These variables were used to perform a principal component analysis (PCA). The PCA was performed in R (v3.6.1, R Core Team, 2019) by the default package STATS, in order to find variables (components) accounting for as much as possible of the variance in environmental data, avoiding collinearity between variables, an expected issue in ecological analysis.

A total of 311 individuals were randomly sampled from the 14 populations (Table 1). Leaves or buds were collected from individual trees and put in silica gel for long-term conservation. Total genomic DNA was extracted by grinding 50–60 mg of dry material in liquid nitrogen using the DNeasy Plant mini kit (QIAGEN). Extracted DNA was quantified spectrophotometrically (Eppendorf Biophotometer^®) and diluted to 5 ng/μl.

Samples were analyzed with a set of eight polymorphic markers (Table 2) developed by Kulju et al. (2004). PCRs were performed in triplex master mixes in Eppendorf Mastercycler^® pro. Each sample was amplified in a 12.5 μl total volume reaction according to the QIAGEN Type-it^® Microsatellite PCR kit. The cycling protocol included: 95°C for 5 min; 28 cycles at 95°C for 30 s, 57°C for 90 s, 72°C for 30 s; 60°C for 30 min. 1 μl of each amplification product was added in a mixture of 20 μl formamide (Applied Biosystems^TM) and 0.3 μl GeneScan^TM 500 LIZ^TM, and denatured at 95°C for 5 min. Samples were run on ABI Prism 3130 Avant DNA sequencer (Applied Biosystems^TM). Results were analyzed using the GeneMapper^® Software v4.1 (Applied Biosystems^TM), and allelic profiles were scored by automatic binning and visual checking.

The PCA of the study sites (Supplementary Figure S2), based on the multivariate analysis of the 15 climatic variables derived in ClimateEU (Supplementary Table S1), separated sites based on their thermal demand along the first component (explaining 83% of the variance), and the precipitation and moisture deficit along the second component (explaining 13.7% of the variance). The autoecology diagram (Figure 2), based on the picture in Beck et al. (2016), showed that ITC6, ITC2, and ITS2 fell outside the climate niche of the Betula species; consequently, they can be considered ecologically marginal. ITS1 and the two Greek sites GR1 and GR2 are at the limit of the climatic niche. ITN3 is shown as an outlier, especially in term of precipitation, but without any limitation in terms of water and thermal demand. The other sites, that include SRB, the Alpine sites ITN1 and ITN2, and the Apennine sites ITC1, ITC3, ITC4, ITC5 can be considered within the envelope of the environmental niche.

A total of 145 alleles was observed across the eight loci (Table 2). The number of alleles per locus varied from 7 (L7.1) to 30 (L1.10), with an average of 18.1. The frequency of null alleles (fna; Table 2) was low, with an average of 2.3% over all loci. The fixation index (F; Table 2) ranged from −0.05 to 0.17, without any significant deviation from zero.

Allelic richness (AR) varied between 3.8 and 7.9 (Table 1). The lowest AR was detected in ITC5 and ITC6, while ITC3, ITN1, ITN2, ITN3, SRB, GR1, and GR2 showed an AR greater than 7.0.

On average, Italian MaP populations showed a lower level of AR compared to Greek MaP populations (5.7 and 7.3, respectively) and to the core populations ITN1, ITN2, ITN3, and SRB (7.4, p < 0.05). The three Balkan populations did not differ in terms of AR (on average 7.2).

The mean number of private alleles (PA; Table 1) ranged between 0.05 in ITC5 and 0.73 in ITC1. Italian Apennine sites showed an average PA equal to 0.35, but this value decreased along the North Apennines - Central Apennines gradient from ITC1 to ITC5 and then it rose again in ITS1 and ITS2. A low value of PA was recorded in Greek populations (on average 0.18). Southern core populations ITN1, ITN2, ITN3, and SRB showed a mean number of PA equal to 0.31.

The level of heterozygosity (HE; Table 1) was on average 0.71. Values ranged from 0.55 (ITC5) to 0.82 (ITN1). Among the Italian MaP populations, a mean HE equal to 0.71 was detected; ITC3 had the highest HE (0.80), while ITC6 and ITC5 presented the lowest values (0.61 and 0.55, respectively). Greek MaP populations (GR1 and GR2) tended to show a higher HE, compared to Italian MaP populations (0.79 vs. 0.71); In addition, their HE was not significantly different from that of the Alpine and SRB populations (both averaged 0.79).

The inbreeding coefficient (FIS; Table 1) ranged between −0.072 (ITC5) and 0.115 (ITC6). FIS for the Italian MaP populations was on average 0.036. A significant departure from FIS = 0 was measured in ITC6. In Greek MaP populations, FIS averaged 0.062. In the Alpine and Serbian populations, FIS was 0.062 and 0.026, respectively.

The power analysis by POWSIM, applied to the 14 populations sampled in this study, with an Fst = 0.11 and 8 SSR markers, resulted in a significant power of 1. When the drift steps were omitted (t = 0), the probability of error when assuming divergence among populations was α = 0.04.

The cluster analysis (DAPC; Figure 3) showed two main clusters formed along the first two coordinates. The first axis grouped the Italian populations of the central Apennines (ITC1, ITC2, ITC3, ITC4, ITC5, and ITC6), while the other embraced the two southernmost Italian sites (ITS1 and ITS2), the Greek populations (GR1and GR2), the Alpine sites (ITN1, ITN2, and ITN3) and the Balkans (SRB). There was further separation along the second axis, especially ITC6 was depicted as clearly isolated from both clusters.

Based on this evident existence of two main clusters, the analysis of molecular variance (AMOVA; Table 3) was performed separately. The cluster ITC1 – ITC2 – ITC3 – ITC4 – ITC5 – ITC6 (Table 3A) showed a highly significant level of differentiation among populations (FST = 0.146, p < 0.05), while diversity within populations (FIS = 0.022) resulted not significant. The other cluster ITS1 – ITS2 – ITN1 – ITN2 – ITN3 – SRB – GR1 – GR2 (Table 3B) presented comparable low values of FIS and FST (0.062 and 0.043, respectively).

The FST pairwise population comparison (Table 4) revealed significant divergence among MaP Italian populations, presenting a mean FST equal to 0.15; on the contrary, GR1 and GR2 were not distinctively differentiated, a result extended to the comparison with SRB. The MaP Italian populations as a whole had an average differentiation when compared to the Alpine populations (ITC1, ITC2, and ITC3) of 0.11.

Considering all the Italian populations (the Alps to Sicily transect), the Mantel test resulted in no correlation between geographic and genetic distances (r = −0.27, p = 0.15). On the contrary, the transect along the Balkans, from northern Greece to SRB, throughout the Alps showed a weak positive significant correlation between geographic and genetic distances (r = 0.38, p < 0.05).

Effective population size (NE) showed large values, with confidence intervals (CI) ranging to infinite in most cases and for all methods (Table 5), without any evidence of genetic drift. The models used to detect population bottlenecks did not show any significant heterozygosity excess averaged across loci in all populations, thus proving the lack of any recent bottleneck event (Table 5).

The STRUCTURE analysis (Figure 4A and Supplementary Figure S1) revealed a highest likelihood partitioning at k = 2 which distinguished between one group that assembled all the Italian northern-central Apennines populations (ITC1, ITC2, ITC3, ITC4, ITC5, and ITC6; referred to as Clu1) and the other that included the southernmost Italian populations, the Alpine and the Balkan populations (ITS1, ITS2, ITN1, ITN2, ITN3, SRB, GR1, and GR2, referred to as Clu2).

Further analysis of the substructure of these two gene pools was performed. Clu1 presented k = 5 as the most likely group partitioning (Figure 4B). The substructure of Clu2 revealed k = 3 as the most likely number of genetic groups, in particular, the ITS1 – ITS2 group, the Alpine ITN1 – ITN2 – ITN3 group and the Balkan SRB – GR1 – GR2 one (Figure 4C). However, some admixture was evident in these groups. When we included only the Italian populations in the structure analysis, a separation into five groups (k = 5) was observed, with the Alpine populations sharing a common gene pool and the Apennines populations more differentiated among each other (Figure 4D).

The Monmonier’s function in ADEGENET recognized genetic barriers around the northern and central Apennine sites (Figure 5). The MSN resulted from EDENnetworks (Figure 5) displayed in the site ITN1 the main node (eight connections), followed by the two Greek sites and SRB (7), ITS1and ITN2 (6), ITS2 (5), and ITN3 (4). Low connections (from 0 to 1) were shown for the nodes in central Apennines populations. Thus, the resulted network (Figure 5) showed strong links among the Alpine sites with the Balkan ones and also, despite the physical distance, among the Alpine sites and the southern Italian sites. Medium strength links were displayed between Southern Italy and the Balkan peninsula. Poor or null connections existed among the northern and central sites in Italy.

Our distinction among the MaP populations with the central populations relied on geographic elements, assuming their distance and peripheral location reflect their ecological requirements, with environmental conditions being optimal at the center and poorer at the periphery (Brown, 1984). The environmental analysis performed in this study partially reflects this view. In fact, all the populations on the Alps and SRB fitted within the environmental niche depicted in the chorological map by Beck et al. (2016), based on datasets of field observations. ITN3 appeared as an outlier similarly to other scattered sites with high precipitation and cold temperatures. The southernmost MaP (ITS2) and the two MaP populations at the lowest elevation (ITC2 and ITC6) were characterized by environmental requirements outside the environmental niche for this species. These three sites all presented dry conditions due to high evaporative demand and elevated temperatures, conditions not typical of birch sites. Also, it is interesting to note that these last three mentioned sites (ITS2, ITC2, and ITC6) are peculiar in their soil substrates. The Greek sites (GR1 and GR2) together with ITC1, are sites located at the border toward dry and warm conditions similarly to ITS2, ITC2, and ITC6. If the future Mediterranean climate is drier and warmer (IPCC, 2013), these populations might face harsher climatic conditions. The other MaP populations of the Apennines (ITC1, ITC3, ITC4, and ITC5), all vegetating at high altitudes, are within the environmental niche, similarly to the Alpine and Serbian populations. Based on this ordination, our classification solely based on geographical elements is not biased, especially for the more elevated sites of the Italian Apennines. This basic view, when integrated with genetics, has to deal with the fact that MaP populations are usually smaller, numerically less abundant and more fragmented when compared to central populations; consequently the reduced effective population size and low gene flow can determine higher genetic differentiation and lower genetic diversity, as a result of stochastic events (genetic drift, low gene flow and excess of inbreeding), as it was the case for the Apennine populations, discussed below.

Our data displayed a high level of genetic diversity at the SSR loci examined in this study, with no sign of excess of inbreeding. Only locus L5.5 tended to show a high FIS, but this resulted not significant, with no alteration of the allelic frequencies from equilibrium. At the population level, the sites on the Italian Apennines (ITC1–ITC2–ITC3–ITC4–ITC5–ITC6–ITS1–ITS2) generally presented a lower level of allelic richness and, to less extent, gene diversity compared to the putative central populations ITN1, ITN2, ITN3, and SRB. Differently from the Italian MaP populations, Greek MaP populations revealed a higher allelic richness and heterozygosity compared to the Italian ones and similar to the Alpine and Serbian populations. This pattern highlights that stochastic events, like genetic drift and low gene flow, might have affected the genetic diversity of the Italian MaP populations, while displays the overall favorable conditions of the Greek populations, in terms of genetic diversity. GR1 and GR2 reflect the other accepted hypothesis about the MaP populations and their level of gene diversity. In fact, according to other authors, marginal populations may harbor the bulk of genetic diversity at the species level (Hewitt, 2004; Hampe and Petit, 2005), as range-wide patterns of genetic diversity are typically shaped by past climate-driven range dynamics than by stochastic events (Eckert et al., 2008). Hence, refugial Greek populations might have been source of migrants during the northward advancement of birch at the end of the last LGM, contributing to the colonization of Balkans and mixing with other central European refugial populations or migrants from the east waves (Petit et al., 2003; Palmé et al., 2003; Birks and Willis, 2010).

Different level of private allelic richness was observed among the Italian MaP populations, with the highest values in ITC1 and ITC2. A lower level of genetic diversity was observed in the southern part of the central Apennines (ITC5) and in ITC6, concomitant with a fewer number of private alleles and allelic richness, while the highest genetic diversity was observed in ITC4, ITS1, and ITS2. The detection of a high level of private alleles, as observed in ITC1, may help to detect putative refugial region; newly colonized areas are characterized by low genetic diversity and lack of private haplotypes, as might be the case for ITC5. However, according to Magri et al. (2015), multiple refugia made up by small isolated scattered populations had been present in these mountains, without any substantial change in extension and distribution. This can also be in accordance with the presence of multiple glaciers (Allen et al., 2008; Adamson et al., 2013) from the northern Apennines to Etna, indicating different spots of cold areas, thus presumably suitable habitats for birch. The ITC6 population, which presented low levels of allelic richness and private alleles, as well as a positive significant inbreeding coefficient, can have been affected by inbreeding among genetically related trees, likely due to selection pressure linked to the characteristic unfavorable environment of this site (low elevation and semi-arid climate, sulfuric water and acid soil). The structure and sub-structure analyses stressed the peculiarity of the site, confirming its genetic divergence. The physiological basis of the population survival under these conditions has not been investigated according to the authors’ knowledge. It is noted that birch fossil pollen has been found in the general area (at 15–24 ka BP), in a site along the Tyrrhenian central coast, inferring a long-standing presence of this species (Tzedakis et al., 2013). Genic selection under a long-lasting environmental pressure has likely led to a unique genetic composition and differentiation in this population. Genome sequencing of natural birch populations revealed that an evident link exists between environmental variable, meanly temperature and precipitation, and candidate gene expression related to cold acclimation and phenological events like cambial activity, wood formation, bud burst, and flowering (Salojärvi et al., 2017). Such genetic change induced by environmental pressure was manifested, for instance, in an intact grassland ecosystem in response to 15 years of simulated climate change (Ravenscroft et al., 2015). It is likely that this phenomenon would happen also in trees, but at different time scales (Kremer et al., 2014; Schierenbeck, 2017).

A slight gradient of diversity along the Apennines at the higher elevation was observed, with a high genetic diversity in the southern Italian MaP populations (ITS1 and ITS2), a lower value in ITC5 and a rise again toward ITC3 and ITC1. In a study on birch genetics by Tsuda et al. (2017), assessed by nuclear SSRs, that sampled Italian populations matching our sites ITN2, ITC1, and ITS1, and that also included a northern Greek population, a similar pattern of gene diversity and allelic richness can be observed, with the northern Apennine population showing a lower level of allelic richness and gene diversity compared to the Alpine, the southern Italian and Greek populations.

Two genetically different groups were distinguished, separating the six MaP populations in central Italy from the two southern Italian populations, which share their gene pool with Alpine, Serbian, and Greek populations. The separation of the six populations of the central Apennines from those in the southern Apennines resembles the result found for silver fir by Piotti et al. (2017). The uniqueness of the northern-central Apennines populations is strengthened by the genetic barriers detected in our Monmonier’s function analysis, showing the lack of gene flow along the Italian peninsula. As some sites are only a few kilometers apart, a mountain island effect and different environmental conditions could have provoked this phenomenon (Schoettle et al., 2012; Yousefzadeh et al., 2016), causing non-overlapping flowering phenology and shaping the pattern of genetic structure even in a species with potentially high gene flow (Ortego et al., 2012). The high statistical significance, resulting when analyzing the power of this structure, made us confident that the sample and marker sizes were sufficient to provide with high probability the depicted genetic population structure. It is often claimed that statistical power is expected to be high when sample and loci sizes are large. These conditions are particularly true at a low level of differentiation and short divergence time (Kalinowski,2005a). The level of Fst = 0.11 detected among these populations can be considered large enough to be caught through our sampling strategy, further supported by the high level of polymorphic loci.

In our study, geographic distances were not correlated to the genetic distances when the Italian populations were tested, which is typical when gene flow between neighboring populations is low, causing differentiation and weakening of the isolation-by-distance effect on the spatial genetic structure. The same result was found for European beech (Fagus sylvatica L.) populations in different European regions (Leonardi et al., 2012). They indicated habitat fragmentation and small population size as the causes of the genetic differentiation pattern observed in marginal beech populations. Although we did not detect any recent bottleneck or genetic drift, the lack of gene flow was evident in the networks analysis. However, especially in anemophilous tree species, gene flow (the product of effective population size and rate of migration among populations) can overtake the negative effects of fragmentation. Such a case is reflected in the Greek populations, which, due to wider extension and more continuous distribution compared to the Italian populations, exhibited a stronger isolation-by-distance along the latitudinal gradient extending toward the Balkan and the Alps.

Our study revealed the existence of historical gene flow between the populations from Southern Apennines and Balkan Peninsula, which is already reported for other tree species. Indeed, studying genetic differentiation and phylogeny of beech on the Balkans employing isoenzyme markers, Gömöry et al. (1999) and Magri et al. (2006) reported that despite nowadays geographical isolation, a similar genetic structure was observed between beech populations from Balkans and southern Italy. Pieces of evidence of common haplotypes, which grouped Italian and Balkan populations, were also found in Turkey oak (Quercus cerris; Bagnoli et al., 2016). Similarly, studies on silver fir (Abies alba) in the Apennines and Balkans found genetic similarity between the southern Apennines and eastern European populations (Belletti et al., 2016; Piotti et al., 2017). These studies are in agreement with ours and point toward the existence of historical gene flow between southern Italy and the Balkan peninsula (Carabeo et al., 2016), when the Apennine and Balkan Peninsulas were connected during Pleistocene at glaciation periods and concomitant sea-level drops (Nieto Feliner, 2014).

Populations along the Apennines in Italy represent examples of MaP populations, characterized by little extension and patchy distribution, low genetic diversity, low gene flow, and high differentiation. In Greece, populations are at the rear edge of the species continuous natural distribution do not show any evidence of genetic erosion. Hence, they represent a case where range dynamics, driven by past climate dynamics, played a major role in shaping the genetic structure. However, their rear edge location is a cause of concern about their preservation, since the rise of the isothermals and the drought spells extension forecasted for the Mediterranean region may endanger their gene pools. Since high levels of genetic variation are expected to increase the potential of the species to respond to selective pressure, we suggest that populations with higher genetic diversity and gene flow, such as the sites in Greece and also in the southern Apennines, can be considered for the establishment of in situ gene conservation units. Concurrently, populations showing peculiar gene pools are also considered as valuable source material for genetic conservation programs. In this regard, the stands of silver birch in the central Apennines and those at the lower elevations are genetically distinct and should also be targeted for conservation efforts, likely ex situ, as their long-term in situ adaptive potential may be compromised. However, genotypes able to survive and reproduce in warmer and drier conditions are expected to be selectively advantaged in local persistence and during migrations at higher latitudes and, particularly in the Mediterranean basin, altitudes. Under this perspective, with ongoing climate changes, these populations can be a source of genetic novelty for introducing and reinforcing standing genetic variation in other parts of the range.