DEEP DIVERSIFICATION AND LONG-TERM PERSISTENCE IN THE SOUTH AMERICAN ‘DRY DIAGONAL’: INTEGRATING CONTINENT-WIDE PHYLOGEOGRAPHY AND DISTRIBUTION MODELING OF GECKOS

SAMPLE COLLECTION

Samples were obtained through fieldwork led by FPW, GRC, or MTR, and through tissue loans obtained from colleagues (Table S1). Vouchers were catalogued in the Coleção Herpetológica da Universidade de Brasília (CHUNB) or Museu de Zoologia da Universidade de São Paulo (MZUSP). Phyllopezus pollicaris is a habitat specialist on rock outcrops, but at regions where these are not abundant, individuals are sometimes associated with human buildings (F. P. Werneck and M. T. Rodrigues, pers. obs.).

The complete dataset for molecular analyses includes 393 tissue samples from 68 localities, spanning the complete geographic, taxonomic, and morphological extent of variation currently known in P. pollicaris. Sampling localities are distributed within and outside modeled historical refugia for P. pollicaris and for the open biomes (Fig. 1). Phyllopezus periosus and P. maranjonensis (Gamble et al. 2012) were used as outgroups in all phylogenetic analyses.

SEQUENCE DATA COLLECTION

We sequenced partial mtDNA sequences from cytochrome b (cytb) and NADH dehydrogenase subunit 2 (ND2) genes from most individuals (some samples failed to amplify for ND2). We then obtained a maximum likelihood (ML) gene tree to direct subsampling efforts representing haplotype and geographical diversity (one or two individuals representing all haplotypes at each locality) for the nuclear loci. We screened approximately 40 rapidly evolving nuclear markers from a variety of sources, and amplified 11 variable loci within P. pollicaris (including protein coding regions and introns). Supporting Information Appendix S2 and Table S2 summarize relevant details for the 13 gene regions, primers, laboratory protocols, sequence assembling, alignment, recombination tests, and treatment of heterozygous individuals. We deposited all sequences in GenBank (accession nos. JQ825288-JQ828670) and trees in TreeBase (no. 12502).

GENE TREE ESTIMATION AND GENETIC DISTANCES

To describe the overall phylogeographic structure of P. pollicaris, we performed ML analyses on single-gene and partitioned concatenated datasets (mtDNA and nuDNA), including both gene copies in the case of phased nuclear loci, using RAxML version 7.0.0 (Stamatakis 2006). Analyses were implemented with GTR + Gamma model, 200 independent ML searches, and 1000 nonparametric bootstrap replicates to assess nodal support (Felsenstein 1985). We used Geophylobuilder version 1.0 (Kidd and Liu 2008) to construct a tridimensional representation (geophylogeny) from the mtDNA gene tree and associated geographical data, and visualized it with ArcGIS and ArcScene version 9.1 (ESRI). We calculated net among-group distances between mtDNA haploclades with MEGA version 5.05 (Tamura et al. 2011), using uncorrected and Tamura–Nei (Tamura and Nei 1993) corrected distances, and 500 bootstrap replicates to estimate standard errors.

SPECIES TREE ESTIMATION AND DIVERGENCE DATING

We estimated the P. pollicaris species tree from the multilocus trees under a coalescent model and simultaneously estimated divergence times using *BEAST version 1.6.2 (Drummond and Rambaut 2007; Heled and Drummond 2010). The lack of Phyllopezus fossils prevented rigorous divergence dating, but we used a less-ideal indirect calibration based on substitution rates to provide a first estimate of divergence dates for the major nodes relevant to testing biogeographic scenarios. We used uncorrelated relaxed clocks to allow for rate heterogeneity among lineages, a Yule prior for the species tree, and a normal prior on the global substitution rate to calibrate the estimation (mean = 0.0065 substitutions/my; SD = 0.0025 for the ucld.mean parameter) based on the mtDNA substitution rate of 0.65% changes/million years (Macey et al. 1998), widely employed in dating squamate phylogenies. For the nuclear loci, we used the default gamma prior for ucld.mean and exponential prior for ucld.stdev, with a mean of 0.5.

As with most species tree estimation methods, *BEAST requires a priori assignment of individual alleles to a species (‘traits’ file) before estimating the relationships. There is no routine procedure on how to assign individuals to species in a poorly known complex, but any approach elected should avoid underestimation of intraspecific diversity. We provisionally made assignments based on well-supported and geographically structured mtDNA haploclades. We selected at random one of the phased nuclear gene copies to represent each individual on this and subsequent analyses to avoid extremely time-consuming computations. We implemented five independent runs of 100 million generations each (total 5 × 10⁸ generations), and assessed convergence of MCMC runs (effective sample sizes, ESS values > 200) using Tracer version 1.4.1 (Rambaut and Drummond 2007). Stationarity was reached before the 10% of the posterior samples, which was used as a conservative burn-in when pooling files from the independent runs with LogCombiner (Drummond and Rambaut 2007). We used BEAGLE version 1.0 (Ayres et al. 2012), a programing interface to accelerate analyses.

POPULATION STRUCTURE AND ASSIGNMENTS

We investigated population structure with a Bayesian probabilistic genetic clustering implemented by STRUCTURE version 2.3 (Pritchard et al. 2000) using a genotype matrix of two mtDNA and four nuDNA. This subset of nuclear loci spanned a range of nucleotide diversities and coalescent times (KIF24, MYH2, PRLR, and rpl35) in the complex, but allowed us to avoid the exceptionally time-consuming computations typical of such large datasets. We are aware that in some instances (e.g. long history of isolation), STRUCTURE can create clusters that are inconsistent with the main evolutionary divisions (Kalinowski 2011). This result seems to be forced by evaluation of an inappropriately small number of discrete genetic clusters, K (Kalinowski 2011), so we explored a large range of values by running 20 replicated analyses over a range of K from 2 to 30. Each of these 580 independent runs implemented one million generations following a burn-in of one million generations, and incorporated the possibility of mixed ancestry. We visualized the optimal K structure based on the rate of change in the log probability of data between successive K values, ΔK (Evanno et al. 2005), as calculated by Structure Harvester.

We combined the replicate analyses under the optimal K identified with CLUMPP (Jakobsson and Rosenberg 2007), and plotted results with individuals in the order of their appearance on the mtDNA gene tree, to facilitate cross-visualization of results. Because ΔK may favor smaller values of K representing basal levels of hierarchical structure (Evanno et al. 2005; Weisrock et al. 2010), we reran the STRUCTURE analyses (same set of parameters) within each of the major basal clusters identified to investigate finer level structure.

QUATERNARY PALAEOMODELING AND ANCESTRAL DISTRIBUTIONS

To estimate P. pollicaris species distribution models (SDMs) across Quaternary climatic fluctuations, we implemented the maximum entropy machine-learning algorithm MAXENT (Phillips et al. 2006). The occurrence dataset included our field GPS records, georeferenced museum or published records, and records from trusted collectors or on-line gazetteers (e.g., ACME Mapper version 2.0, and Global Gazetteer version 2.2), which were verified with Google Earth to ensure points were not located in heavily urbanized areas. Current climatic variables were downloaded from WorldClim (Hijmans et al. 2005), while Last Glacial Maximum [21 ky BP, LGM] and Holocene [6 ky BP] were based on ECHAM3 (Deutsches 1992), and Last Interglacial [130 ky BP, LIG] data were obtained from Otto-Bliesner et al. (2006). We then obtained a historical stability map by overlapping the presence/absence projections under the four scenarios. Methodology is described in detail in Werneck et al. (2011; in press), as applied to the open biomes. To estimate the ancestral distribution of the basal node of P. pollicaris, we used PhyloMapper version 1b1 (Lemmon and Lemmon 2008) with nonparametric rate-smoothing to obtain an ultrametric tree from the mtDNA gene tree based on unique haplotypes, and corresponding geographical coordinates.

GENETIC DIVERSITY AND TEST OF QUATERNARY PHYLOGEOGRAPHIC PREDICTIONS

We calculated population genetic metrics and tested general genetic predictions derived from the habitat stability (refugia) hypothesis (Table 1) based on mtDNA (Carnaval et al. 2009). We grouped samples based on the results of Structure, and into stable or unstable regions, and calculated coalescent estimators of genetic diversity with Migrate-n version 3.2.15 (Beerli and Palczewski 2010). We used two independent ML runs with 10 short chains and five long chains, for a total of 2000 and 20,000 generations, a burn-in of 20,000 steps, and starting parameters for the calculations derived from an F_st-like estimator (Beerli and Palczewski 2010). We calculated the other genetic parameters in Table 1 and respective significance levels with 10,000 coalescent simulations using DnaSP version 5 (Rozas et al. 2003). We also implemented Mantel tests with Zt (Bonnett and de Peer 2002) to investigate isolation-by-distance (IBD) patterns as characterized by correlations between geographical and genetic distances matrixes.

TESTS OF POPULATION DIVERGENCE AND BIOGEOGRAPHIC SCENARIOS

We used an ABC approach to test three alternative scenarios for the diversification of P. pollicaris relevant to hypothesized biogeographic scenarios (models and parameters described in Results). Analyses consisted of three steps: (1) draw parameter values from the prior distributions and simulate data under each model; (2) compute SuSt and the distance between simulated and observed datasets; and (3) use a rejection algorithm to approximate the posterior distribution of parameters by retaining the simulations closest to the observed data, based on a specified threshold, that is tolerance (Lopes and Beaumont 2010). Model choice can affect estimates and conclusions, so alternative models should represent biologically relevant hypotheses in the simplest way possible, and model fit should be explored (Carstens and Knowles 2010; Csilléry et al. 2010). We simulated 100,000 coalescent genealogies under each model with msABC (Pavlidis et al. 2010), and simulated a total of six loci (the same used for Structure analyses) (Pavlidis et al. 2010). Number of gene copies, number and length of loci, and prior distributions were chosen to reproduce observed values from the empirical dataset.

We estimated empirical relative mutation rates of loci and average theta across lineages specified in the models using Migrate-n version 3.2.15 (Beerli and Palczewski 2010) implementing the same search strategy as above, and used those parameters to convert absolute times into coalescent times (units of 4N_o generations), assuming one generation per year. We selected all seven global SuSt to be computed with mean and variance calculated across the loci, but excluded population-specific SuSt that cannot be compared between models with different numbers of populations. We transformed observed sequence data into ms-like files and then calculated the same global SuSt across all loci.

We estimated posterior probabilities and models support using ‘postpr’ and parameter posterior distributions using ‘abc’ functions of the R package ‘abc’, move the comma to out of the single quote implemented in R version 2.11.1 (R Development Core Team 2010). We set tolerance to 0.001 and implemented a nonlinear neural network regression, which has been shown to outperform other ABC algorithms (Camargo et al. 2012). We summarized simulated model fit to the observed data using a Principal Components Analysis (PCA): predictive plots in which the observed SuSt occurred within the cloud of simulated SuSt were interpreted as good fit. All intensive computational analyses were submitted to the BYU Fulton Supercomputing Cluster.

MARKER AND DNA POLYMORPHISM

We sequenced a total of 1828 and 7477 bp of mtDNA and nuclear markers, respectively. Nucleotide diversity ranged from 0.59% (RBMX) to 16.85% (ND2); and overall haplotype diversity and average number of nucleotide differences were surprisingly high (Table 2). We found a high number of polymorphic sites for the mtDNA and most nuclear loci, including some protein coding genes (Table 2). The two mtDNA fragments were concatenated and treated as a single locus in subsequent analyses; 269 of the 388 samples analyzed represented unique haplotypes restricted to single localities. Most well-sampled localities (n≥ 4) were characterized by different haplotypes, but haplotypes were shared once between two nearby localities (Xingó[no. 12] and Poço Redondo [no. 16]), and once between two disjunct populations (Tianguá[no. 20] and Peixe [no. 39]).

All but three nuclear loci (MC1R, MAP1B, DMXL1) had at least one fixed indel among lineages, and MYH2 and rpl35 regions had the largest number of indels. Barra do Garças (no. 57) and São Geraldo do Araguaia (no. 23) had the highest frequency of fixed indels (across five loci), followed by Babaçulândia (no. 24), Grão Mogol (no. 51), and Mucugê (no. 36), with indels across three loci.

GENE TREE ESTIMATION

The mtDNA tree identified multiple deep, moderately to strongly supported haploclades, with several localities exhibiting monophyly. We identified 44 mtDNA haploclades (A–AR, Fig. 2, Table S1), which were used as assignments for species tree estimation. We conservatively recognized some haploclades with low support, if concordant with geography, to avoid underestimating diversity in downstream analysis. Corrected mtDNA distances among haploclades ranged from 0.2% between geographically close sister-clades (Table S3) to values as high as 35% (uncorrected distances 0.2–25%, Table S4), with mean distance across the tree of 23.4% (SD = 0.9%).

Some individual nuclear gene trees lacked resolution at the most basal levels (e.g., ACA), but most had moderate resolution and support (Fig. S1). In contrast, the concatenated nuclear ML tree recovered 12 strongly supported clades (Fig. 3). The overall genealogy of mtDNA and nuclear trees was congruent, with some relationships consistently recovered and well supported across all trees, as P. p. przewalskii (from southwestern Brazil and throughout Chaco) and the Caatinga group (Figs. 2, 3, and S1). Cerrado populations occupy three positions on the tree, two small basal lineages, including mtDNA haploclades A, B, and C, and nuclear clades (Mucugê no. 36) and (São Geraldo do Araguaia no. 23 + Barra do Garças no. 57), sister to all other lineages. A second clade, mtDNA haploclade E and nuclear clade (Serra da Mesa no. 44, Lajeado no. 37, and Palmas no. 38), is sister to the strongly supported clade comprising all P. p. przewalskii populations (Fig. 3). The third group (mtDNA haploclades Q–U and V–W) is closely related to a clade including most Caatinga localities (haploclades X–AR). With a few exceptions, populations from the Caatinga formed a well-supported clade deeply nested within the tree, but with internal relationships poorly resolved and short branches.

SPECIES TREE ESTIMATION AND DIVERGENCE DATING

We observed large ESS (> 200) and convergence for most parameters across independent and combined runs for the species tree analysis. Divergence time estimates were bound by broad confidence intervals (95% highest posterior density, HPD) and, in many instances, were not calculated at nodes of ambiguous relationships, that is, posterior probabilities lower than 0.5 (Fig. S2); thus, these are provisional conservative estimates. The species tree based on all 13 loci (Fig. 4) is mostly concordant in topology with both mtDNA and nuclear concantenated ML trees. As in previous analyses, São Geraldo do Araguaia no. 23, Barra do Garças no. 57, and Mucugê no. 36 (clades I and II) are sister to all other P. pollicaris lineages, and diverged in the mid-Miocene (∼11.5 Ma). The large clade of populations from Caatinga is deeply nested in the species tree and diverged in the Mid–Late Miocene, but internal support is low (Fig. 4, green). The strongly supported P. p. przewalskii clade was isolated approximately 8.5 Ma, and shows moderate to strong support for internal relationships (Fig. 4, blue).

Divergence times estimated the origin and major crown clades splits of P. pollicaris to have occurred during the Miocene (between 11 and 5 Ma), followed by divergence within these major groups during the Pliocene–Pleistocene transition (5–1.5 Ma). These estimates are concordant with previous estimates for the origin of P. pollicaris (Gamble et al. 2011), and indicate that limited diversification occurred during Late Pleistocene, mostly among geographically close populations in the Chaco and Caatinga (Fig. 4, blue and green). Most Cerrado populations initiated and ended their diversification during the Miocene, and are characterized by long branches and older diversification dates (Fig. 4, red).

POPULATION STRUCTURE AND ASSIGNMENTS

Calculations of ΔK detected a peak at K= 3, representing the most basal hierarchical structure in the data (Evanno et al. 2005; Weisrock et al. 2010). Assignments detected at K= 3 are congruent with geographic clusters in the Southwest (SW), Central (CE), and Northeast (NE) of the P. pollicaris distribution, which are largely coincident with the distributions of the Chaco, Cerrado, and Caatinga, respectively (Figs. 5,6). One exception are samples from a few close localities in Paraíba state (Itaporanga [no. 7], Junco do Seridó[no. 4], Patos [no. 5]), assigned to the Central cluster, while surrounded by populations assigned to the Northeast cluster (Fig. 5). There is also overall concordance between identified multilocus clusters and mtDNA haploclades and geophylogeny (Fig. 5). Some individuals located at the contact regions between clusters revealed admixed assignments (Fig. 6).

When we reran analyses within each of the major clusters, ΔK detected K= 3, K= 9, and K= 3 (total K= 15) for the SW, CE, and NE clusters, respectively, demonstrating further population structure, more pronounced in the CE populations. However, although most individual membership posterior probabilities were greater than 0.9 in the first round of analyses, coefficients are overall lower at the finer level structure, and further divide both admixed and nonadmixed sets of individuals (Fig. 2). Therefore, as we considered using population assignments to develop divergence models, and alternative models should preferentially represent biologically relevant questions in the simplest way possible, we chose to focus subsequent analyses at the basal level of clustering (K= 3).

PALAEOMODELING AND ANCESTRAL DISTRIBUTIONS

Estimated palaeomodels depict a potential distribution of P. pollicaris contracted to the southwest and northeast of the ‘dry diagonal’ during the LGM, with a break in central Brazil. These two SDMs blocks closely track the distribution of population clusters, SW to one extreme and [CE + NE] to the other (Fig. 6). However, when considering all SDMs and the stability surface, we observe that P. pollicaris potential distribution did not change drastically across Quaternary climatic fluctuations, especially from the Holocene to the present (Fig. 6). Note the extensive overlap of highly suitable areas across the four time periods (i.e., large areas of stability; Fig. 6, yellow areas). Despite minor oscillations in the species range, the P. pollicaris stability map provides a close match for the ‘dry diagonal’ biomes stability models (Werneck 2011; Werneck et al. in press). Thus, we are confident that P. pollicaris tracked changes of the open biomes and that this complex represents a suitable proxy of their biogeographic patterns at both shallow and deep temporal scales. The ancestral distribution was estimated by PhyloMapper to be located in central Brazil, nearby CE and NE current distributions, where ancient Phyllopezus lineages became extinct (Fig. 6, black circle).

GENETIC POLYMORPHISM AND TEST OF PHYLOGEOGRAPHIC PREDICTIONS

None of the groupings (or the total sample) revealed detectable signatures of population expansion or reduction (Table 3). We did detect significant IBD across all groupings, except for the SW cluster, which had overall lower diversity when compared to the other clusters (Table 3). We did not find major differences between samples from stable and unstable areas with respect to most population genetic metrics, contradicting predictions from the hypothesis of strong influence of Quaternary refugia on clade's genetic structure (Hewitt 2004; Carnaval et al. 2009). In fact, unstable areas presented higher genetic diversity than stable areas with respect to several metrics (Table 3).

TESTS OF POPULATION DIVERGENCE AND BIOGEOGRAPHIC SCENARIOS: ABC

We developed three models representing alternative biogeographic scenarios for the diversification of P. pollicaris across the ‘dry diagonal’ based on (1) speciation expectations from estimated genealogies, population structure, and palaeomodels; and (2) available geological information with relevant times that fall within the dated chronogram from *BEAST. The use of external information (e.g., dates of geological events) is crucial to narrow the universe of possible histories (Garrick et al. 2010). The first model (A) represents a no-speciation null model and reflects views from early studies that the open biomes had a shared history (Vanzolini 1976); it was modeled as a large population, with no substructure and free migration. Model B hypothesizes an approximately simultaneous early divergence from the ancestral area (as estimated by PhyloMapper) into three clades (SW, CE, and NE, as recovered by Structure), and represents a geologically older speciation scenario. Model C represents an early diversification of the southwest (SW) group from the other groups (CE + NE), followed by a more recent divergence between these last two. This hypothesis would correspond to two speciation events within P. pollicaris, and it predicts a (SW [CE + NE]) topology (Fig. 7).

We attribute the early divergence time (T1) to the final uplift of the Central Brazilian Plateau that took place 7–5 Ma, an important event hypothesized to have driven divergence across the open biomes (Werneck 2011). We hypothesize the more recent divergence time (T2) to represent the final phase of the ecological expansion of grass-dominated vegetation and increased fire frequency (5–3 Mya), which may have promoted additional ecological diversification of the Cerrado from the adjacent open biomes (Beerling and Osborne 2006; Simon et al. 2009; Edwards et al. 2010). Both divergence times fall within estimated divergence times for P. pollicaris lineages (Fig. 4). Although the use of three subgroups estimated by Structure does not fully corroborate the phylogenetic results (which recover a paraphyletic CE group), the models were intended to capture the broad substructure and biogeographic patterns supported by available geological data for the ‘dry diagonal’. Because divergence with some degree of gene flow may be common (Nosil 2008), we also considered the possibility of migration in the diversification models B and C. Model A included the per locus mutation parameter θ (= 4Noμ) for a single lineage across multiple loci, whereas models B and C included additional parameters: the divergence times in coalescent units (τ, in units of 4N_o generations) and the migration rate between divergent lineages (m, prior varied from 0, no migration, to 1, high migration). We assumed stable population sizes in all models.

When compared simultaneously, ABC could not attain resolution between the three models, unless the tolerance rate was increased (Table 4). ABC favored the speciation scenarios over the no-speciation scenario in all pairwise comparisons. More specifically, diversification of P. pollicaris into two large groups, SW and CE + NE (model C), was selected in every comparison that included this model (Table 4). Posterior probabilities were overall lower for comparisons that included migration, relative to equivalent comparisons without migration, reflecting increased difficulty in distinguishing between speciation and no-speciation models when migration is present (Nielsen and Wakeley 2001). Bayes factors of pairwise comparisons to the preferred models attained moderate values (up to ∼12). The observed SuSt occurred within the bounds of the prior sample of simulated SuSt at PCA predictive plots, confirming good model fitting (Fig. S3).

DNA POLYMORPHISM, GENETIC DISTANCES, AND POPULATION STRUCTURE

Patterns of DNA polymorphism, gene and species trees are clearly consistent with the occurrence of a complex of cryptic species within P. pollicaris, as hypothesized by Gamble et al. (2011, 2012). The large mtDNA genetic distances among clades found here exceed typical values ranging from 5% to 11% reported among gecko sister species (Rocha et al. 2009; Fujita et al. 2010), and for gecko complexes with likewise high hidden genetic diversity and low haplotype sharing between localities (Geurgas and Rodrigues 2010).

The P. pollicaris species complex has a strong population structure in which the three major clusters at the deepest levels are concordant with the geographic limits of the Chaco, Cerrado, and Caatinga biomes; and also approximately congruent in the gene and species trees and geophylogeny, implying that these ‘dry diagonal’ biomes do not share a single diversification history. Although the SW cluster is more isolated from the others, the CE and NE are intermingled over several regions of overlap, as revealed by admixture assignments and lower diversity in the SW Chaco cluster. The CE cluster is characterized by higher population subdivision, genetic diversity, and nucleotide differences across localities than the other two clusters. This cluster is mostly confined to the Cerrado, which has a landscape compartmentalized between ancient plateaus and young valleys (Cole 1986). This heterogeneity at local and regional scales potentially enhances environmental opportunity for ecological divergence and speciation (Werneck 2011), if these changes result in reduced gene flow and isolation between lineages (Rundell and Price 2009). Geomorphological complexity and heterogeneity are then likely to have shaped the genetic structure among CE populations.

MAJOR RELATIONSHIPS AND DIVERGENCE TIMES

Analyses of mtDNA and nuclear data showed overall congruence in the major relationships, differing mostly in their branch support at deeper levels (higher for most nuclear loci). We found higher support for the nuclear concatenated analysis relative to species trees, which have been suggested to be an artifact of uninformative loci for short branches, and still comprise a challenge for species-tree methods for higher level phylogenies (Townsend et al. 2011). As uninformative nuclear loci are more frequent at the lower levels, we expected such an artifact to have stronger influences on species-trees estimated from phylogeographic datasets. Indeed, McCormack et al. (2011) also reported slightly lower topological support for the species tree in Aphelocoma jays, although their study included only three nuclear markers.

Instances of topological discordance are evident at internal nodes of more recent lineages, and at the admixture zones between CE and NE phylogeographic groups. For example, relationships within the NE group are poorly resolved. This pattern can be attributed to the reduced topographic complexity of the Caatinga that is disrupted by few prominent landforms, such as elevated plateaus (e.g., Chapadas do Araripe, Apodi, and Borborema, Serra do Espinhaço), and relictual rainforest enclaves (Sampaio 1995). Fewer geological barriers in the Caatinga likely enhanced opportunity for gene flow, yielding less-resolved diversification histories.

A few samples are characterized by discordant placement among datasets or methods. For example, the clade D (Diamantina no. 55 + Serro no. 56) occupying a basal position in the mtDNA (Fig. 2) and the species trees (Fig. 4), but a more nested position in the nuclear concatenated tree (Fig. 3). Similarly, Matias Cardoso no. 50 + Augusto de Lima no. 54 and Grão Mogol no. 51 (clades V and W) are basal in the species tree but deeply nested in both concatenated trees, and these samples show admixed structures (Fig. 6). These cases show the confounding effect of contact zones between adjacent groups on species tree inference, and the importance of multilocus coalescence methods for estimating species trees in instances of discordance (Leaché 2009). In these cases, we favor the species tree hypothesis that places some of the northern Minas Gerais populations closer to the base of the tree (V, W, and D), and remaining samples (Januária [Q, no. 49] and Manga [R, no. 48]) as sister to the large clade of Caatinga populations (Fig. 4).

The most interesting outcome of this hypothesis comes from the disjunct placement of Manga (clade R, no. 48) and Matias Cardoso (clade V, no. 50) localities that are less than 15 km apart, but are separated by the São Francisco River (Fig. 6, letter a, SFR). The SFR is recognized as a barrier that has driven speciation at high taxonomic levels, with several endemic genera and species pairs isolated on opposite sides (Rodrigues 1996). More recently, the SFR was suggested as an important phylogeographic barrier to Atlantic Forest taxa distributed closer to the delta (Pellegrino et al. 2005; Carnaval and Moritz 2008). Recent divergence estimates in the lizard genera Eurolophosaurus (Passoni et al. 2008) and Calyptommatus (Siedchlag et al. 2010) suggest that the SFR probably formed a vicariant barrier during the Late Miocene, concordant with P. pollicaris divergence times. The SFR also likely prevented secondary contact between previously diverged P. pollicaris lineages, reinforcing its importance as a singular center of diversification. Our results also suggest a role for the Tocantins River as barrier between CE and NE clusters (Fig. 6, letter b). Denser sampling and investigation of other taxa distributed in the region are, however, needed to further explore this hypothesis.

São Geraldo do Araguaia (no. 23/A), Barra do Garças (no. 57/B), and Mucugê (no. 36/C) on the other hand were recovered as highly distinct and early differentiated populations in all analyses. Mucugê together with other populations (Diamantina [no. 55/D] and Serro [no. 56/D]) are located in the high altitude “campos rupestres” in the Serra do Espinhaço mountain range (Fig. 6, letter c), a region well known for its high levels of endemism, including squamates (Cassimiro and Rodrigues 2009; Nogueira et al. 2011). Establishment of the Espinhaço range was a regional event that potentially influenced P. pollicaris divergence (see below).

Divergence within the P. pollicaris complex proceeded in a nearly continuous fashion since about 11.5 Ma, but with crown clades originating in the Mid–Late Miocene. Phyllopezus pollicaris przewalskii from Chaco is strongly supported by gene and species tree, and diverged approximately 8.5 Ma. Pleistocene diversification is recovered among some geographically close Caatinga populations; a time correlating with geological evidence for the exposition of formerly submerged granite surfaces (lajedos) where the species occur. These times suggest a major influence of divergence/geographical barriers during the Miocene, followed by a nearly continuous diversification, with no clear bursts of Quaternary diversifications (Rull 2011b). Long-term diversification with combined Neogene and Pleistocene causal mechanisms is expected to be apparent at higher taxonomic levels (Rull 2011a), whereas the prolonged complex history of ‘dry diagonal’ biomes is here evident within a single complex.

Likewise, ancient divergence dates have been reported in other gecko complexes (Pellegrino et al. 2005; Geurgas and Rodrigues 2010; Oliver et al. 2010), and these are commonly older than codistributed taxonomic groups. Patterns of New World gecko diversification exemplify most of the scenarios described for terrestrial vertebrates, including ancient vicariance, trans-Atlantic rafting and temporary land bridge dispersal, and human introductions (Gamble et al. 2011), and reflect their ancient Gondwanan origin (Kluge 1987). Positive correlations between clade age and species diversity described for geckos (Gamble et al. 2011) can account for some of the patterns we report for P. pollicaris complex (ancient divergence times and high diversity).

MODEL-BASED POPULATION DIVERGENCE: BIOGEOGRAPHIC IMPLICATIONS

We are aware that our models represent simplifications (e.g., no substructure within major clusters, use of combined splitting times, and use of a nearly monophyletic group [CE]), but they do reflect phylogeographic patterns within P. pollicaris that correlate with major biogeographic events across the ‘dry diagonal’. ABC favored the speciation scenarios over the no-speciation model, and we found stronger support for model C, representing an early diversification (∼7–5 Ma) of the southwest (SW) from the (CE + NE) group, followed by a more recent divergence between these last two (∼5–3 Ma). Comparisons of model C under isolation and migration suggest that divergence may have occurred with some gene flow. Due to the high correspondence between groups tested and the ‘dry diagonal’ biomes, we can infer geomorphological events potentially responsible for an early diversification of Chaco lineages followed by divergence of Cerrado and Caatinga lineages, and their biogeographic implications.

Key Neogene events often emphasized as causal mechanisms for Neotropical diversification (e.g., closure of the Panama Isthmus, Andean uplift, change of drainage patterns with the cessation of marine incursions into the Amazon basin; Hoorn et al. 2010; Rull 2011a) almost certainly impacted indirectly the ‘dry diagonal’ biota diversification, which is largely confined to the Brazilian Shield (Werneck 2011). However, other correlated, but less-understood, events were more relevant for this region and for the diversification of P. pollicaris complex. From the hypothesized ancestral area in central Brazil, some ancestral population dispersed across the ‘dry diagonal’ before being subject of the following suggested biogeographic events that generated the three major genetic lineages.

Marine transgressions during the Mid–Late Miocene (∼10–5 Ma), supported by sedimentology and marine fossil records, covered large portions of the Chaco-Paraná Basin (CPB), and formed the second flooding pulse of the Paranaense Sea, isolating large areas of the Chaco depression and associated biota (Hernández et al. 2005; Ruskin et al. 2011). Isolation of southwest populations of P. pollicaris initiated by these marine transgressions was most likely enhanced by the Chaco subsidence due to an intense uplift phase of the Andes and the final uplift of the Central Brazilian Shield during the Late Miocene–Early Pliocene transition, 7–5 Mya (Gubbles et al. 1993; Uba et al. 2006; Mulch et al. 2010). We hypothesize that these events, in concert, drove allopatric speciation of the P. p. przewalskii clade (SW). The resulting altitudinal break acted as a geographical barrier and prevented secondary contact, with the lowland populations isolated on the SW side (CPB and Paraguay River basin) and the higher elevation Brazilian Plateaus populations to the CE–NE side (Fig. 6). Strong isolation coupled with low densities of P. pollicaris preferred habitat (rock outcrops) in the Chaco, may have limited population sizes, and thus explain the reduced molecular variation in the SW clade.

We also hypothesize that subsequent regional tectonic events responsible for the final establishment of the Brazilian Plateau and enclosed regional mountain ranges (e.g., Serras do Espinhaço, do Mar, da Mantiqueira, 2–4 Mya) coupled with ecological events such as expansion of grasslands and increased fire frequency (Beerling and Osborne 2006; Simon et al. 2009) would promote differentiation between Cerrado and Caatinga clades. However, because CE and NE lineages are not reciprocally monophyletic and show some admixture, they may have diverged through ecological speciation along environmental gradients, without physical barriers to prevent gene flow (Nosil 2008). The alternative of sympatric speciation is supported by an increasing body of evidence in marine vertebrates (Foote et al. 2009; Crow et al. 2010), where obvious barriers are less evident, but is also documented for terrestrial vertebrates (Niemiller et al. 2008; VanderWerf et al. 2010), and must also be considered as a possibility here.

Although the exact mechanisms are not clear, events at the Tertiary–Quaternary boundary that significantly remodeled eastern Brazil's landscapes may have confounded the genetic signatures of the earlier isolating events, including vegetation shifts (Werneck et al. 2011, in press), and neotectonic activity (Riccomini and Assumpção 1999). We suggest that the less dramatic differentiation between Cerrado and Caatinga lineages could be due to a combination of more recent divergence and gene flow, and represent a possible example of ecological or sympatric speciation (Vanzolini 1976).

PAST DISTRIBUTIONS AND TESTS OF QUATERNARY PHYLOGEOGRAPHIC PREDICTIONS

Our results suggest a limited influence for Quaternary climatic fluctuations on the phylogeographic history and SDMs of P. pollicaris. Patterns of genetic variation and structure reflect the impact of earlier influences and a Miocene origin of the crown clades followed by persistence through the Pleistocene without major demographic changes. Our modeling suggests a slight range reduction of P. pollicaris during the LGM with a break in central Brazil. This “two LGM refugia” scenario reflects very distinct SDMs between the SW and the [CE + NE] clusters, and these ecological niche differences may have promoted additional diversification between these two groups.

Although Pleistocene distributional shifts shaped strong genetic signatures in many groups (Carnaval et al. 2009; Knowles and Alvarado-Serrano 2010; Lessa et al. 2010), there are counter-examples in which phylogeographic signatures reflect events that have not been overwritten by Pleistocene climate dynamics (Leavitt et al. 2007; Bell et al. 2010; Thomé et al. 2010; Hoskin et al. 2011). For P. pollicaris specifically, the influence of climatic fluctuations seems minor when compared to geographical and topographical influences that remained largely unchanged since the Miocene, however, they can not be neglected, and likely played a relevant role to promote further differentiation between and within previously diverged lineages.

TAXONOMIC IMPLICATIONS AND PROVISIONAL SPECIES LIMITS

Phyllopezus pollicaris should clearly be treated as polytypic, and because the consistently recovered clades are deeply divergent and strongly supported in all analyses, they are almost certainly valid species. Here, we comment briefly on species limits and taxonomic implications within P. pollicaris complex.

The P. p. przewalskii clade is strongly supported across all analyses, and it has a distinct structure and a singular SDM. Although morphological uniformity generally characterizes the P. pollicaris complex, P. p. przewalskii is also morphologically diagnosable by the lack of postcloacal tubercles and slightly lower ventral scales count (Vanzolini 1953), and in its karyotype (Pellegrino et al. 1997). Under a general lineage species concept (de Queiroz 2007) and our empirical evidence for both ecological and genetic nonexchangeability (Crandall et al. 2000), P. p. przewalskii should be elevated to species status. The nested placement of P. p. przewalskii within P. pollicaris renders the other subspecies, P. p. pollicaris, paraphyletic. We hypothesize that the following clades (Roman numerals in Fig. 4 and Table S1) should be treated as ‘candidate species’: Clades I (no. 38) and II (no. 7 and no. 54), which are the sister groups to all other P. pollicaris, should be treated as two candidate species, along with these six remaining clades: Clade III (6 localities); Clade IV (3 localities); Clade V (P. przewaslkii, 11 localities); Clade VI (6 localities); Clade VII (7 localities), and the Clade VIII (the 32 remaining Caatinga localities). See Figure S4 for the geographic distribution of the eight ‘candidate species’. Aside from P. przewaslkii, we do not suggest formal names for any of these lineages, but as ‘candidate species’ (Morando et al. 2003) they can guide biogeography-based conservation planning because they register broad evolutionary patterns without the necessity of formal names (Bickford et al. 2007; Riddle et al. 2011).

IMPLICATIONS FOR ‘DRY DIAGONAL’ CONSERVATION

Recent studies show that current knowledge of amphibian and reptile species diversity in the Neotropics has been underestimated (Fouquet et al. 2007; Geurgas and Rodrigues 2010; Gamble et al. 2011; Funk et al. 2012). The cryptic genetic diversity disclosed here reveals that the South American ‘dry diagonal’ biomes are no exception. With the management task of preserving adaptive diversity, evolutionary processes, and natural genetic connections across the range of species complexes (Crandall et al. 2000; Davis et al. 2008), we can propose some conservation implications of our results.

Our assessment of the P. pollicaris complex shows that protection of the eight candidate species is necessary to preserve the evolutionary processes that generated them, especially if these geographic regions have fostered codivergence in other taxa (Davis et al. 2008). Protection of long-isolated populations should be augmented by protection of admixture zones to maximize adaptive diversity and evolutionary potential (Crandall et al. 2000; Smith et al. 2001; Moritz et al. 2009). This can be accomplished by placing reserves at the core of the three major clusters, and several are established and were included in our sampling (e.g., no. 6, 10, 16, 19, 40, 42, 48, 49, 50, 64; Fig. 1). Fortunately, reserves have also been established in the transition zones between CE and NE clusters (no. 23 and 26), but additional reserves in this region (e.g., no. 36, 37, 38, 51, 54, 55, 56,) and between the SW populations and the other lineages (e.g. no. 57, 58, 61) should be considered in regional conservation planning.

Persistence through Quaternary fluctuations and high diversity within the P. pollicaris complex suggests that it has the potential to tolerate adverse scenarios and to adapt to new conditions (Davis et al. 2005), as long as rock outcrops persist. These patterns provide important insights about responses to future environmental changes and long-term population viability, a critical variable for establishing efficient conservation strategies (Diniz-Filho et al. 2008). However, other taxa associated with ‘dry diagonal’ features more susceptible to oscillations may be more susceptible to extinction due to climate change. Allocation of conservation resources will then be more effective if comparative studies can provide evolutionary histories of a diverse array of codistributed ‘dry diagonal’ endemics.