Necessary permits for fieldwork and sampling were obtained from the Corporación Nacional Forestal (CONAF-Chile).

Fresh leaves from a total 558 individuals from 17 populations of D. humilis were collected throughout its entire distribution range. Fifteen populations (Dhh01-Dhh15) corresponded to D. humilis subsp. humilis and two populations (Dhp01-Dhp02) to D. humilis subsp. polyanthes (Table 1, Fig. 1). Eight populations of D. humilis subsp. humilis (Dhh01-Dhh08) were located North of the Maipo river basin, growing in semi-arid Mediterranean-type climate areas, whereas the other seven populations (Dhh09-Dhh15) were located South of it (Table 1, Fig. 1). Five of these (Dhh09-Dhh14) were growing in sub-humid Mediterranean-type climate areas, and the southernmost population (Dhh15), together with two populations of D. humilis subsp. polyanthes (Dhp01-Dhp02), were growing in humid Mediterranean-type climate areas [20]. DNA extraction followed the procedure described in [21]. Individuals were genotyped for eight unlinked microsatellite following [22].

Two plastid regions, trnT-L and trnL-F [23] were amplified and sequenced in up to six individuals per population following [21]. Sequences were deposited in Genbank under the accession numbers KF357945-KF357955. A combined matrix of individual sequences of both plastid regions totalling 54 sequences was used in subsequent analyses.

Allele frequencies and genetic diversity indices were calculated in all populations using GENETIX 4.05 [24]. Deviations from Hardy-Weinberg equilibrium were tested in all populations using GENEPOP v. 4.0 [25]. Different taxonomic and geographical population groups were compared to reveal differences in average values of allelic richness (A*), observed heterozygosity (H_O), genetic diversity within populations (H_S), inbreeding coefficient (F_IS) and population differentiation (F_ST) using FSTAT v. 2.9.3.2 [26], and differences were tested for significance with 10,000-permutation tests. Population pairwise differentiation (F_ST) was calculated with ARLEQUIN 3.11 [27] and tested for significance using 1000 replicates. ARLEQUIN was also used to generate a matrix of pairwise linearized F_ST values (i.e. F_ST/(1−F_ST); [28]), which was correlated to a log-transformed matrix of geographical distances between populations to test for Isolation By Distance (IBD) through Mantel tests. Significance of correlation was tested with 1000 permutations with NTSYSpc 2.11 [29].

Pairwise D_A genetic distances [30] between populations were calculated with POPULATIONS 1.2.3 [31] and used to conduct a Principal Coordinates Analysis (PCO) and a Minimum Spanning Tree (MST) that was superimposed on the PCO plots using NTSYSpc 2.11.

Population genetic structure was investigated by means of Analysis of Molecular Variance (AMOVA) which was performed in ARLEQUIN 3.11 to partition in different population groups according to the taxonomical or geographical membership. The significance of the analyses was tested with 1000 replicates.

Bayesian clustering was also used to infer population genetic structure using STRUCTURE 2.1 [32]. Analyses were based on an admixture ancestry model with correlated allele frequencies, for a range of K genetic clusters from one to 19, with ten replicates for each K. The analyses were performed with a burn-in period and a run length of the Monte Carlo Markov Chain (MCMC) of 7×10⁵ and 7×10⁶ iterations, respectively. The most likely number of genetic clusters (K) was determined according to Evanno et al. [33].

Haplotype polymorphism was estimated within populations and within genetic and geographical groups through the analysis of the number of segregating sites (S), the number of haplotypes (h), the haplotype diversity index (Hd) and the average number of pairwise nucleotide differences between DNA sequences (θπ) [34] with DnaSP5 [35]. Indels encompassing two to several nucleotides were reduced to single gaps and treated as a fifth nucleotide state for a statistical parsimony haplotype network analysis with TCS v. 1.21 software [36].

Environmental niche modelling (ENM) was conducted to evaluate the potential distribution of the geographical groups of D. humilis under current climatic conditions and under Last Glacial Maximum (LGM) and Last Interglacial (LIG) conditions. A set of 19 bioclimatic variables (Table S1 in Appendix S1) retrieved from WorldClim (www.worldclim.org) plus the altitude were used, and GIS layers with 30 sec resolution were clipped to the extent of central Chilean regions using DIVA-GIS [37]. Correlation among environmental variables was determined by Mantel tests using XLSTAT and tested for significance with 1000 random permutations (Table S1 in Appendix S1). Then we selected a reduced set of nine uncorrelated environmental variables with higher percent contribution (PC) and permutation importance (PI) based on jackknife pseudosampling on the ENM of D. humilis (Tables S1 and S2 in Appendix S1): altitude, bio3 (isothermality), bio4 (temperature seasonality), bio6 (minimum temperature of coldest month), bio7 (annual range temperature), bio9 (mean temperature of driest quarter), bio15 (precipitation seasonality), bio18 (precipitation of warmest quarter) and bio19 (precipitation of coldest quarter).

Additionally, we assessed pairwise correlations between all 20 environmental variables studied and pairwise D_A population genetic distances [30], and pairwise population linearized F_ST [28], among populations of D. humilis, and the correlation between the 20 environmental variables and latitude, using the Mantel test with 1000 random permutations.

The maximum entropy algorithm implemented in MAXENT v. 3.3.3k [38]–[39] was used to construct the models. Maxent is optimal for ENM when using small sample sizes [40] and when environmental predictions are poorly influenced by the addition of irrelevant bioclimatic variables [41]. The D. humilis data fit these requirements since no significant increase in the area under the curve (AUC) values was observed when using all variables compared to those from the reduced set of variables (Table S3 in Appendix S1).

Occurrence data were split into training data (75%) to build the model and test data (25%) to test the accuracy of the model. Fifteen subsample replicates were performed in each run using the default options and 1000 iterations. Model accuracy was assessed with the AUC value of the receiver-operating characteristic curve (ROC) [38]. The contribution of each environmental variable to the ENM was evaluated through a Jackknife pseudosampling (see above). A tenth percentile threshold was applied for all models.

ENM were conducted for the two northern (Dhh01-Dhh08) and southern (Dhh09-Dhh15) population groups of D. humilis ssp. humilis. The low number of known populations of D. humilis subsp. polyanthes (Dhp01-02) precluded a confident ENM analysis of this taxon.

ENMs were projected to LGM (c. 21 ka BP), with 2.5 arc-minutes resolution [42], and to LIG (c. 120–140 ka BP), with 30 arc-seconds resolution [43] scenarios. Two palaeoclimatic layers simulated for two general atmospheric circulation models were used for LGM: the Community Climate System Model (CCSM, [44]) and the Model for Interdisciplinary Research on Climate (MIROC, [45]). Both CCSM and MIROC layers were combined following a conservative approach by including their overlapping predicted areas [46]. Current minimum predicted values were used to determine the past minimal predicted areas, assuming that the environmental requirements of D. humilis subsp. humilis have remained stable during at least since LIG.

A complementary ENM approach was done through a Principal Component Analysis (PCA), which was constructed with the raw data obtained from the 19 climatic variables and the altitude for each population of D. humilis using PAST 2.17c [47], Fig. S1 in Appendix S1).

All eight microsatellite loci were polymorphic and amplified a total of 79 alleles in the 17 studied populations of D. humilis (Table S4 in Appendix S2). The number of alleles per locus ranged from three (B633) to 24 (H442) with a mean of 9.88±6.38 (±SD) alleles per locus. The mean number of alleles per locus and population ranged from 3.25 (Dhh12) to 5.00 (Dhh11, Table 1). Of the 79 microsatellite alleles scored, 34 (43.04%) were shared by both subspecies of D. humilis, while 36 (45.57%) and 9 (11.39%) were exclusive to D. humilis subsp. humilis and D. humilis subsp. polyanthes, respectively (Table S4 in Appendix S2).

Observed heterozygosities ranged from 0.350 (Dhh04) to 0.598 (Dhp01), and unbiased expected heterozygosities from 0.357 (Dhh12) to 0.521 (Dhh10) (Table 1). Five of the 17 populations showed HW deviations towards heterozygote deficiency; three, including one population of D. humilis subsp. polyanthes, showed non-significant departure from HW equilibrium, and the remaining eight populations of D. humilis subsp. humilis and one of D. humilis subsp. polyanthes showed a significant heterozygote excess (Table 1).

No significant differences were detected for the tested genetic diversity indices between D. humilis subsp. humilis and D. humilis subsp. polyanthes, except for observed heterozygosity (H_o). Surprisingly, the more restricted endemic D. humilis subsp. polyanthes showed significantly higher (p = 0.033) average H_o (Table 1). Similarly, the comparison of northern and southern population groups of D. humilis subsp. humilis failed to find significant differences at any of the tested indices (Table 1).

Moderate but significant (different from zero; p<0.05) levels of population differentiation were observed among populations (results not shown). Higher average F_ST values were found between populations of both subspecies (average F_ST = 0.295) than among populations within subspecies (average Dhh F_ST = 0.145; Dhp F_ST = 0.011). Similarly, a higher average differentiation was observed between northern and southern populations groups of D. humilis subsp. humilis (average F_ST = 0.198), than among populations within northern populations (Dhh01-Dhh08 F_ST = 0.069) and southern populations (average Dhh09-Dhh15 F_ST = 0.109) of D. humilis subsp. humilis with differentiation between the groups not being statistically significant (Table 2).

Bayesian analysis of population structure showed a maximum ΔK = 1598.45 value for K = 3 (Fig. S2 in Appendix S2). In this clustering, individuals of D. humilis subsp. polyanthes showed a high proportion of membership to cluster 3 and those of D. humilis subsp. humilis to cluster 1 (populations Dhh01-Dhh08) or to cluster 2 (populations Dhh09-Dhh15; Fig. 1a). Mean F_ST values corresponding to the divergence between clusters 1, 2 and 3 and the hypothetical ancestral population were 0.114, 0.201 and 0.266, respectively, indicating that populations showing a higher membership to cluster 1 were less diverged from the ancestral population. A further maximum ΔK = 127.90 value was obtained for K = 5 (Fig. S2 in Appendix S2) which separated the populations of D. humilis subsp. humilis into two additional genetic clusters (clusters 1–4; Fig. 1b).

Non-hierarchical AMOVA attributed 19.03% of the total variation to among populations of D. humilis s.l., and 15.00% of the total variation to among populations of D. humilis subsp. humilis (Table 2). In hierarchical AMOVA, the largest proportion of variation among groups (21.37%) was obtained for a taxonomical grouping of populations into subspecies. AMOVA based on a geographical grouping of populations attributed 12.15% of the variation to differences between northern and southern groups of D. humilis subsp. humilis and a lower proportion of variance (7.65%) to differences among populations within groups (Table 3). The grouping of D. humilis subsp. humilis populations into four genetic clusters did not increase the variance among groups (11.81%) but lowered the proportion of variance among populations within groups (5.34%).

PCO showed results consistent with STRUCTURE (Figs. 1 and 2). Populations of D. humilis subsp. polyanthes separated at a large distance from populations of D. humilis subsp. humilis (Fig. 2). Clustering of populations of this latter taxon was consistent with their geographical distribution (Fig. 2). PCO with superimposed MST analysis identified the closer relationship of D. humilis subsp. polyanthes to the southern populations of D. humilis subsp. humilis (Fig. 2).

A significant correlation between pairwise geographical distances and linearized F_ST values was found in both D. humilis s.l. (r = 0.537, p = 0.001) and D. humilis subsp. humilis (r = 0.416, p = 0.004) populations (Fig. 3a), thus showing significant isolation by distance (IBD). However, the pattern of IBD vanished when this analysis was separately conducted within both northern (r = 0.296, p = 0.080) and southern (r = 0.005, p = 0.420) geographical groups of D. humilis subp. humilis (Fig. 3b).

The combination of trnL–F and trnT–L plastid DNA regions produced eight haplotypes (Fig. 4; Table S5 in Appendix S2). Six and one haplotypes were restricted to D. humilis subsp. humilis and subsp. polyanthes respectively, and one haplotype was shared between both taxa. Haplotype IV was widespread in 12 populations of D. humilis, including one population of D. humilis subsp. polyanthes, and had the highest outgroup probability (0.771). Three haplotypes (I, II and III) were restricted to some western populations of northern D. humilis subsp. humilis, whereas haplotypes V, VI and VII were restricted to some southern populations of D. humilis subsp. humilis (Fig. 4a; Table S5 in Appendix S2).

Haplotype diversity was higher in Dioscorea humilis subsp. humilis (S = 6, θπ = 0.789), than in D. humilis subsp. polyanthes (S = 1, θπ = 0.500) (Table 3), as expected for the wider distribution range and population abundance of the former. The northern group of populations of D. humilis subsp. humilis (Dhh01-Dhh08) showed higher haplotype diversity (S = 3, θπ = 0.873) compared to the southern group (Dhh09-Dhh15) of populations (S = 3, θπ = 0.534) (Table 3).

TCS estimated a 95% maximum connection of 17 steps incorporating all eight haplotypes into the network and inferred three missing haplotypes (Fig. 4b). The haplotype network showed a star-like pattern with four of the six derived haplotypes directly connected to the most widespread one (Hap. IV; Fig. 4). Three haplotypes were connected to the central one at a larger number of mutations. Two of them were private to the northernmost population (Hap. I and II) whereas the other (Hap. V) was private to a population from the southern group (Fig. 4).

All rainfall-derived variables (bio12-bio19) and all but three temperature-derived variables (bio6, bio8 and bio11) were significantly correlated to genetic distances (Table S1 in Appendix S1). Also, latitude was highly correlated to all rainfall-derived variables (bio12-bio19), to two temperature-derived variables (bio2-bio3), and to both pairwise populations’ D_A and F_ST genetic distances (Table S1 in Appendix S1).

According to response curves and Jackknife tests, the most informative variables for the ENM of D. humilis s.l. were altitude and three climatic variables derived from rainfall data (bio16, precipitation wettest quarter; bio18, precipitation warmest quarter; and bio19, precipitation coldest quarter). At the subspecies level, the variables with the largest contributions to the ENM of D. humilis subsp. humilis were altitude and the climatic variables bio8 (mean temperature of wettest quarter), bio9 (mean temperature of driest quarter) and bio15 (precipitation seasonality) (Table S2 in Appendix S1). Independent ENM for northern (Dhh01-Dhh08) and southern (Dhh09-Dhh15) genetic groups of D. humilis subsp. humilis revealed that the variables bio15 and bio18 were most informative for the northern group, whereas bio8, bio9 and bio15 were most informative for the southern group (Table S2 in Appendix S1). All projections (Figs. 5a–5c) showed excellent predictive success rates, with AUC values higher than 0.9 (Table S3 in Appendix S1). The PCA of environmental variables for the D. humilis populations (Fig. S1 in Appendix S1) separated northern from southern genetic groups of D. humilis subsp. humilis. The southernmost population, Dhh15, showed a distinct set of climatic conditions from the others, according to its separated position, and clustered together with the D. humilis subsp. polyanthes populations in the bidimensional PCA plot (Fig. S1 in Appendix S1).

The ENM for current environmental conditions was mostly concordant with the current distribution of northern and southern genetic groups of D. humilis subsp. humilis (Fig. 5c), except for the southernmost population, Dhh15. Potential areas of contact were predicted on the eastern boundary of their distributions. Projections to the LIG (Fig. 5a) predicted a minimal extension area for the potential distribution of the southern group and a larger extension for the northern group. Projections to the LGM (Fig. 5b) predicted a substantial reduction in the southern group and, to a lesser extent, of the northern group. The present and the two historical models predicted areas with unsuitable environmental conditions between the northern and southern groups (Fig. 5), and an increase in the potential areas of contact between the two groups since the LGM to the present.

Genetic diversity and population structure in plant species is determined by various abiotic and biotic factors, some of which have triggered population differentiation [48] and speciation processes [49]–[50]. Biotic factors have been globally assigned to life-history (e.g. life form), and reproductive traits (e.g. reproductive systems, pollination and seed dispersal mechanisms [51]). However, closely related taxa, such as the two subspecies of D. humilis, show similar biotic parameters. Abiotic factors include, as most relevant, climatic variables and barriers to dispersal. Our results suggest that the conjunction of these two later factors caused the intraspecific split within D. humilis.

Our analyses revealed moderate levels of allelic diversity and heterozygosity across populations of D. humilis (A = 3.25–5.0, H_O = 0.350–0.598, H_E = 0.357–521, Table 1), which were relatively lower than in the sister species D. biloba (A = 5.14–7.29, H_O = 0.345–0.686, H_E = 0.458–0.706, Table S6 in Appendix S3). However, genetic diversity parameters of D. humilis were in the range of other yam species with a different combination of life-history (climbers), reproductive (winged seeds) and distribution range (broad range, non-endemic), characteristics that, contrary to D. humilis should predispose them to higher levels of genetic diversity (Table S6 in Appendix S3). By contrast, D. humilis showed higher genetic diversity than those of species of the Borderea group of Dioscorea which are comparable in morphological and reproductive traits (Table S6 in Appendix S3), though the Pyrenean Dioscorea species differ from D. humilis in their even narrower distributions [52]–[53].

Widespread taxa tend to maintain higher levels of genetic diversity compared to geographically restricted congeners [54]. However, genetic diversity in D. humilis subsp. polyanthes was not significantly lower than in D. humilis subsp. humilis, despite its more restricted geographical range (Table 1). This result could suggest equally or more efficient mechanisms buffering against genetic loss in D. humilis subsp. polyanthes.

Our study also revealed a strong geographical structure of nuclear microsatellite variation throughout the range of D. humilis (Fig. 1), with populations of D. humilis subsp. polyanthes clearly separate from those of D. humilis subsp. humilis that split into clusters 1 (northern populations Dhh01-Dhh08) and 2 (southern populations Dhh09-Dhh15; Fig. 1a). Clustering analyses (Fig. 2) and AMOVA (Table 3) also found a major differentiation between the two subspecies, in agreement with their morphological distinction [16]. However, plastid DNA haplotype sharing (haplotype IV) between the subspecies and the occurrence of a private plastid haplotype (VIII) in Dhp01, directly derived from the most common haplotype (IV), suggests recent diversification mediated by isolation by distance (Fig. 3a), with incomplete lineage sorting in D. humilis subsp. polyanthes (Fig. 4) or alternatively, introgression between subspecies.

Bayesian F_ST values supported an origin of the species in the northern region and a derived recent origin of D. humilis subsp. polyanthes from the southern group of D. humilis subsp. humilis. This was corroborated by the highest haplotypic diversity of the northern group (Table 3) and by the PCO-MST analysis, which indicated the closeness of D. humilis subsp. polyanthes to the southern D. humilis subsp. humilis populations (Fig. 2).

The intraspecific divergence of the southern D. humilis subsp. polyanthes from the southern group of D. humilis subsp. humilis could have been a consequence of local environmental adaptation. However, despite the lack of an ENM for the former taxon, the range of values for its 19 bioclimatic variables overlap with that of the latter group and are not significantly different from them (Table S2 in Appendix S1). Thus, the explanation for their divergence is other than a climatically driven speciation process; it may be rather the consequence of geographical isolation and incomplete plastid lineage sorting of D. humilis subsp. polyanthes from the central Chilean depression southern D. humilis subsp. humilis group during the last glacial and interglacial phases (Fig. 5).

A noticeable finding was the strong geographical structure detected among D. humilis subsp. humilis populations, which separated into two North-to-South genetic groups (Fig. 1a). Such spatial patterns are usually driven by the effect of strong geographical or climatic barriers to dispersal, as proposed for Hordeum chilense Roem. & Schult., which is similarly distributed along a climatic gradient in Chile [20]. The distribution of the genetic groups of D. humilis subsp. humilis mostly paralleled those of the main Chilean latitudinal climatic zones (Fig. 1). Mountain chains in the central Chilean depression show lower altitudes and may not significantly contribute to latitudinal isolation of populations. Our study showed that the genetic divergence of the two population groups occurred northwards and southwards of the Maipo river basin (Fig. 1). Indeed, river basins have been identified as efficient barriers to dispersal for seed plants, such as in Chinese populations of Vitex negundo L. (Verbenaceae) on opposite shores of the Yangtze river [55]. Specifically, the role of the Maipo river as a geographical barrier has been highlighted for other organisms with potentially higher dispersal capabilities than D. humilis, such as the snake Philodryas chamissonis Wiegmann [14]. The Maipo river acting as a geographical barrier to gene flow could contribute to explain the IBD pattern across the range of D. humilis subsp. humilis (Fig. 3a), and the abrupt difference in genetic structure between the two geographical groups (Fig. 1a). However, this IBD pattern does not apply for within-group pairwise population comparisons (Fig. 3b). The absence of IBD within the two geographical areas of D. humilis subsp. humilis contrasts with life-history and reproductive traits of the species which all point towards extremely short dispersal distances [16]. Therefore, the observed patterns are probably mirroring historical gene flow among populations within ranges preceding range expansions in the Holocene and a relatively rapid postglacial expansion by unknown vectors.

Geographical and historical variations of environmental variables have been demonstrated to greatly influence genetic divergence among populations [20], [56]. Our ENM analyses indicate a strong latitudinal ecological differentiation throughout the current range of D. humilis subsp. humilis into two well defined environmental niches (Fig. 5c). Past projection of niche models indicate that this ecological niche differentiation likely originated earlier than LIG (Fig. 5a), and that ecological conditions have been maintained until present times. The current separation of the groups by the Maipo river basin [14] matches the ecogeographical division of the D. humilis range into northern semiarid and southern subhumid Mediterranean areas (Fig. 1, [20]). Ice-cover during the LGM, which reached to 35°S in the Andes [5][6]), together with its northwards influence that extended to approximately 33°S, could have strengthened the barrier effect of the Maipo river. Water volume of this river originating from the Andes was likely higher during LGM than in present times [14], which could account for the allopatric distribution of the two population groups, thereby contributing to the observed genetic differentiation between them (Fig. 5). Bayesian F_ST values supported the ancestry of the northern populations, suggesting a likely origin of the species in its northern range, in the overlapping area with its congener D. biloba [16], followed by further southwards expansion.

Nonetheless, the predicted extension of the potential distribution areas of the two population groups could have fluctuated both in latitude and longitude during glacial and interglacial episodes, as expected from changes in climatic parameters in those areas following periods of warming (LIG and present) and cooling (LGM, Fig. 5). The predicted distribution area of the northern population group showed a maximum extension during LIG (Fig. 5a), whereas the strong reduction during LGM (Fig. 5b) was maintained until present times (Fig. 5c). Contrastingly, the predicted distribution area of the southern group showed a progressive increase in extension from LIG (Fig. 5a) to present times (Fig. 5c). The potential overlap of distribution areas between the two groups reached its maximum extension during present times (Fig. 5c). However, it was restricted to the eastern range of the present distribution of the species, suggesting that lineage migration and admixture, as denoted by the occurrence of the common plastid haplotype IV, is likely to have occurred only along the eastern boundaries of both distribution areas (Fig. 5). Predicted environmental niche models of the northern and southern groups of D. humilis were also consistent with a contraction towards the central Chilean depression during the LGM (Fig. 5b), preceded by broader eastern and western distributions of the potential areas of the northern group during the LIG (Fig. 5a).

Bayesian analyses of nuclear microsatellite variation and of plastid haplotypes also revealed genetic signatures of postglacial population expansion within the northern and southern groups of D. humilis subsp. humilis (Figs. 1b, 4a). Concerning the northern group, western populations predominantly showed a microsatellite genetic membership to cluster 1, whereas eastern populations showed a predominant genetic membership to cluster 2 (Fig. 1a). This was paralleled to a lesser extent by the slower-evolving cpDNA data (Fig. 4a), where three northwestern populations showed three cpDNA haplotypes that were not represented in eastern populations (Fig. 4a). This would indicate a further isolation of the northwestern populations, which, unlike the eastern ones, did not admix with the southern ones.

Contrastingly, a North-to-South expansion was detected in the southern group of D. humilis subsp. humilis, supported by a gradual North-to-South decrease in microsatellite genetic membership to cluster 4, and an increase in membership to cluster 3 (Fig. 1b), agreeing with the predicted postglacial southwards expansion (Fig. 5c). Exclusive cpDNA haplotypes were scattered among populations in this range, and were all derived directly from the most common haplotype (Fig. 4b), suggesting recent divergence and dispersal [57]–[58].