The concatenated cytb and cox2 sequences from the 48 field samples (Table 1) amounted to 1824 bp and contained no indels. There were 468 variable sites defining 29 haplotypes. A fragment of 246 bp was obtained from museum samples of 16 individuals (Table 2). Among these 16 samples, there were 48 variable sites defining 9 haplotypes. 24 Cheirogaleus haplotypes from GenBank, consisting of 17 complete and five partial (307–933 bp) cytb sequences, and two complete and one partial (529 bp) cox2 sequences (Table 3), were aligned with the field and museum haplotypes resulting in an overall set of 62 haplotypes defined through 494 variable sites (Table 4; all sampling sites are given in Fig. 1).

Nuclear DNA sequence data were generated from the 48 field samples. In all individuals both alleles were scored, amounting to a data set of 96 sequences for each locus. Among Cheirogaleus samples, the 370 bp exonic adora3 fragment had 26 variable sites and 29 haplotypes. The 604 bp intronic fragment fiba had 44 variable sites and 49 haplotypes. The 793 bp intronic fragment vWF had 93 variable sites and 52 haplotypes (Table 4). The adora3 alignment contained no indels. Both the fiba and vWF alignments contained a small number of 1–2 bp indels. In addition, the vWF alignment contained indels of 19 and 242 bp in six and three individuals, respectively.

The cytb, cox2, and adora3 loci were each found to best fit a general time-reversible (GTR) model according to AIC. The mtDNA loci were best fit to a model with a proportion of invariant sites (I) and gamma distributed rate heterogeneity (Γ), whereas adora3 was best fit to a model with a proportion of invariant sites. A K81uf+I+Γ model was favored for the fiba locus, (analyzed under a GTR+I+Γ model in Bayesian phylogenetic analyses). The vWF locus was found to best fit an HKY+I+Γ model (Table 4).

Bayesian and ML analyses of the mtDNA data set resulted in congruent trees with three main clades (A, B and C in Fig. 2) that largely correspond to the three species recognized prior to the taxonomic revisions of Groves 24. Clade A is strongly supported (ML BP = 100 and Bayesian PP = 1.0) and consists of haplotypes sampled from western Madagascar, the southeastern tip (Fort Dauphin region) and two sampling sites in the northeast. All mtDNA sequences generated from museum samples of C. medius, C. adipicaudatus, and C. sibreei are placed in clade A.

Clade B is strongly supported (ML BP = 99 and Bayesian PP = 1.0) and is comprised of mtDNA haplotypes sampled from localities along the east coast from the southeastern tip (Fort Dauphin region) to the Maroantsetra peninsula in the northeast (localities 11 and 12). All mtDNA sequences generated from C. major and C. ravus samples fall into clade B. MtDNA sequence generated from a museum sample of C. crossleyi (locality 22) is also placed in Clade B.

Clade C is strongly supported (ML BP = 96 and Bayesian PP = 1.0) and also contains eastern-sampled haplotypes, ranging from the southeastern tip (Fort Dauphin region) up to the northern tip (Montagne d'Ambre, locality 6). Clade C also contains haplotypes sampled from three localities in the northwestern portion of the island (localities 1, 3, and 39). The sole museum-generated sequence placed in Clade C is the C. crossleyi individual from locality 14.

The mtDNA-based relationships among clades A, B and C are poorly supported and are best viewed as unresolved. Uncorrected "p" distances based on the cytb locus (1140 bp) were calculated for the three main clades (A, B and C). Pairwise distances between the three clades were fairly similar, with an average 11.4% between clades B and C, 12.6% between clades A and B and 13.10% between clades A and C.

Bayesian, ML, and statistical parsimony analyses of the individual nuclear loci resulted in generally congruent gene trees with respect to the resolution of clades A, B, and C identified in the mtDNA gene tree. Bayesian and ML hierarchical nuclear gene trees are presented as figures in additional files 1, 2, 3: vWF, fiba and adora3 ML phylogram. Statistical parsimony haplotype networks are presented in Figs. 3, 4, 5. The adora3 haplotype network (Fig. 3) resolves a clade of haplotypes corresponding to clade C in the mtDNA gene tree. The remaining adora3 haplotypes collectively correspond to clades A and B in the mtDNA gene tree. Shared polymorphism of adora3 haplotypes exists among some individuals assigned to these two mtDNA-based clades (Fig. 3). Adora3 haplotype 5 is found in individuals sampled from Ambanja (3), Ambato (4), and Kirindy (37) in the west (all containing Clade A mtDNA haplotypes), and is also sampled from Andrambovato (25) in the east (containing a Clade B mtDNA haplotype). Adora3 haplotype 2 is found in individuals sampled from Bemaraha (38) in the west (mtDNA Clade A) and in the Andrambovato locality (mtDNA Clade B).

The fiba haplotype network consists of two terminal clades that correspond to clades A and C in the mtDNA gene tree. An internal clade is also resolved corresponding to mtDNA clade B. All haplotypes in this latter clade have a common ancestor in fiba haplotype 8 found in a number of individuals sampled from Marolambo (21). These three clades are shallowly diverged from each other. Only two mutational steps separate sampled haplotypes in clades A and B. Only four mutations separate clades B and C.

The vWF haplotype network consists of three clades of haplotypes that nearly completely correspond to clades A, B, and C in the mtDNA gene tree. The sole exception to this pattern is individual RMR149 from Andrambovato, which has a clade B mtDNA haplotype, but is homozygous for a "clade A" vWF allele.

Bayesian population structure analyses of a combined mtDNA and nuclear data set and a data set comprised of only nuclear loci reveal very similar results (additional file 4: Bayesian population structure analysis), indicating that genetic structuring results are not being driven solely by the mtDNA data. Overall, differences between the results of the two data sets were only found in the number of identical solutions found for each K across replicates, in the exact contribution of each K to the genetic makeup of an individual and in the order that individuals split off to from a separate cluster at K = 4. At K > 6 the number of identical solutions plummets to 0 at a 95% threshold. A K = 6 is the favored solution according to the estimated ln probability of the data (mtDNA + nDNA: average = -1033.2, stdev = 2.4; nDNA: average = -862.0, stdev = 9.5), and according to the ad-hoc statistic ΔK 40, which detected a clear mode at K = 6 for the calculations based on four loci, but showed no clear signal for the three nuclear loci. The K = 6 results from analyses of the combined nuclear data set are described below in the context of the three main mtDNA clades (Fig. 6).

Most individuals possessing clade A mtDNA haplotypes are placed with high PPs in two distinct population clusters (depicted in green and purple in Fig. 6). The genetic compositions of individuals from Ambanja/Ambato (locality 4), Kirindy (37), and Ambanja/Benavony (3) are almost entirely of a single population cluster (green). A subset of individuals from Sambava (9) and the single individual sampled from Bekaraoka (7) are placed almost entirely in a second distinct population cluster (purple). Together with two individuals from Bemaraha (38), which do not fall into either of these two population clusters, the individuals forming these two distinct clusters correspond to clade A mtDNA haplotypes.

All individuals containing clade C mtDNA haplotypes are comprised of two population genetic clusters (depicted in blue and orange in Fig. 6). The remaining Sambava individuals are either completely comprised of, or contain high proportions of, a third population genetic cluster (orange) and low proportions of a fourth population genetic cluster (blue). This pattern is reversed in individuals sampled from Andrambovato/Oranjasty (25), Montagne d'Ambre (6), Ambanja/Beandroana (3), Ankazomivady (23) and Manantenina (10).

Most individuals possessing Clade B mtDNA haplotypes constitute two population genetic clusters (depicted in red and yellow in Fig. 6). All individuals from Marolambo and two individuals from Tampolo are either completely comprised of, or contain some proportion of, a fifth population genetic cluster (red). The individuals sampled from Ivorona, Manantantely, Andrambovato/Ambalavero and the remaining individual from Tampolo, are placed entirely in a sixth cluster (yellow). The sole exception is one of the individuals from Andrambovato/Ambalavero, which is only placed in this cluster with a very low PP.

There is a clear indication that many individuals within mtDNA clades contain a mixed nuclear genetic composition. For example, more than half of individuals with clade C mtDNA haplotypes exhibit a genetic composition from two nuclear-defined clusters (orange and blue).

This pattern also extends across mtDNA-defined clades. For example, some individuals from mtDNA clades A and B can contain a high proportion of a nuclear genetic cluster (blue) that is predominantly found in individuals with clade C mtDNA haplotypes. Overall, these patterns demonstrate the existence of two distinct nuclear genetic clusters within each mtDNA-based clade, but demonstrate the potential for extensive shared genetic makeup within and among these clades.

Clade A is strongly supported in ML and Bayesian analyses of mtDNA sequence data and is also supported by genealogical and population genetic analysis of nuclear data, although shared polymorphism of adora3 and vWF haplotypes among mtDNA-based clades A and B were detected. These patterns of shared polymorphism may be the result of a low mutation rate, incomplete lineage sorting, or hybridization. Nonetheless, these patterns are limited and do not obscure the concordant patterns of divergence seen across all loci.

The genetically-derived Clade A clearly corresponds to C. medius sensu lato according to morphometric and geographic data, as well as the mtDNA sequences generated from C. medius museum samples. Analyses of C. adipicaudatus museum specimens invariably place individuals into the C. medius morphometric cluster and mtDNA clade. These results strongly suggest that C. adipicaudatus is synonymous with C. medius, with no evidence indicating that it is divergent from other populations of C. medius.

The C. sibreei individual included in this study is also placed into the C. medius clade A. In fact, it shares its mtDNA haplotype with individual RMR162, an individual most likely belonging to C. medius according to morphometric characters. These two individuals also cluster together in morphometric analyses (LF Groeneveld, unpublished data). As with C. adipicaudatus, these results do not lend support to the hypothesis of C. sibreei as a distinct evolutionary group.

Clade B is strongly supported in ML and Bayesian analyses of mtDNA and corroborated by nuclear data, although with the patterns of shared polymorphism described above. According to analyses of morphological and mtDNA sequence data generated from museum specimens, clade B corresponds to C. major. Analyses of morphological and mtDNA sequence data of C. ravus specimens are congruent and place these individuals into the C. major morphometric cluster and mtDNA clade B. There is no evidence to suggest that C. ravus represents a distinct genetic lineage within the larger C. major group and we conclude that C. ravus is synonymous with C. major. Interestingly, one of the museum-sampled individuals of C. crossleyi (Mu1053) contains an mtDNA haplotype that is placed within mtDNA clade B. However, it is difficult to infer anything else about this C. crossleyi museum specimen since this is a juvenile individual and was not included in the morphological analyses. Other putatively C. crossleyi museum and field-collected individuals are consistently placed outside the C. major group, suggesting that the taxonomic designation of the Mu1053 museum individual may be incorrect.

Clade C is well supported in ML and Bayesian mtDNA gene trees and is unambiguously corroborated by all three nuclear markers. This group contains individuals identified as C. crossleyi in a previous mtDNA-based study 25 and also a single museum individual (Mu1050). In addition to the mtDNA-based identification of this group, morphometric analyses find similarity between the new field-sampled individuals and the Mu1050 museum C. crossleyi individual (LF Groeneveld, unpublished data).

It is important to note here a slight pattern of incongruence between the morphological and molecular results, with respect to C. crossleyi and C. major. Two newly sampled individuals, RMR146 and RMR164, are classified as C. crossleyi individuals according to all genetic data analyses, but cluster with C. major individuals in morphometric analyses. One of these individuals (RMR146) is sampled from a locality (25) that contains both species, suggesting that hybrid introgression may in part be responsible for such a pattern.

We have no molecular data to directly assess the status of C. minusculus. The single specimen upon which this species was described is listed in the museum catalogue of the British Museum of Natural History (NHM) as being from Ambositra/Antsirabe 45. Our geographically closest sampling site to this locality is Ankazomivady (23), which is about 29 km south of the town of Ambositra. Morphometric and genetic analyses indicate these are C. crossleyi individuals, and they do not cluster with the C. minusculus individual in morphometric analyses. Thus, there is no indication that these individuals could represent C. minusculus. Unfortunately, the NHM does not allow sampling of holotype material for molecular analyses, which would be crucial to assess the status of this proposed species. A discussion within the museum community and a change in policy regarding invasive sampling of holotype material for molecular studies is needed to solve this dilemma for future studies.

Within the three species-level evolutionary lineages further genetic substructuring was detected via population genetic clustering. Genetic substructure was most tightly correlated with geography among populations within the clade A lineage, where some populations from the west and northwest (3, 4, 37) were placed in a distinct nuclear cluster, while populations from the northeast (7, 9) were placed in a second distinct cluster. These patterns indicate the potential for the further geographic isolation and divergence of populations beyond the three main Cheirogaleus lineages. However, the lack of concordant geographic patterns of divergence in the gene trees, especially in the mtDNA gene tree, indicate that this divergence may be very recent in nature. It is important to note, however, that Clade A is not entirely comprised of two genetic clusters, but also contains individuals that have cluster assignments more similar to individuals in the other two lineages. This same pattern is seen among individuals from Clades B and C, despite the fact that none of these individuals share alleles. It is possible that this pattern may result from an insufficient level of variation to accurately assign individuals. Additional molecular sampling in future studies is likely to be important in the adequate assessment of population-level structuring using genotypic clustering methods.

Our results do indicate, however, the special status of the Sambava locality; two species, C. medius and C. crossleyi, are found at this locality, and for both species the Sambava individuals are assigned to a genotypic cluster that largely separates them from other individuals of the respective species. This is most strongly seen in the case of C. medius and also includes the single sampled individual from the geographically close locality of Bekaraoka. Posterior probabilities for cluster assignments for some C. crossleyi Sambava individuals are low, but many individuals have high posterior probabilities (≥ 0.90) that clearly distinguish them from other populations of C. crossleyi. Again, these patterns appear to be signatures of more recent episodes of geographic isolation and divergence within the three main lineages of Cheirogaleus.

Results from the new sampling sites identified by this study allow us to further clarify the distribution of the three species of Cheirogaleus (Fig. 1). All three species are present in forest fragments in the Fort Dauphin region (southeastern Madagascar), as described by Hapke et al. 25. The range of C. medius extends along the west coast up north to Ankarana (5). This species is, however, also found at two sites, Bekaraoka (7) and Sambava (9), on the northeastern coast. This is consistent with a presumed C. medius population at Daraina in northeastern Madagascar, as listed in Mittermeier et al. 46. The distribution of C. medius is therefore not limited strictly to the western dry forests. The range of C. crossleyi extends from the southeastern tip of the island all the way north to Montagne d'Ambre (6) with many sampling sites being found inland along the eastern edge of the central plateau. There are a few exceptions in the north, both on the east and west coast: Sambava (9), Iharana/Vohemar (8), Ambanja/Beandroana (3), Manongarivo (1) and Ampijoroa (39). Cheirogaleus major is found, as previously described, in the eastern lowland forest, from the southeastern tip as far north as Maroantsetra (11). Additionally, there is one field (25) and one museum sampling site (22) along the eastern edge of the central plateau where C. major is found. We currently do not have any data to assess the populations mentioned in Thalmann and Rakotoarison 44, Ausilio and Raveloanrinoro 41, Thalmann 43 and Rasolofoson et al. 42.

This study has shown that multiple independent lines of data can yield a robust estimate of species diversity. Morphological data alone are not expected to provide resolution of previously cryptic species, but indeed have had the opposite effect of over estimating species diversity 47. The sole use of mtDNA might have diagnosed distinct populations, or even local matrilines as species and the use of any single nuclear marker would most likely lead to an underestimate of diversity. Only through combining all of these sources of data were we able to achieve a robust estimate of lineage diversity in Cheirogaleus and detect geographically structured patterns of genetic variation. Of course, it would be desirable to add more geographic and ecological information to the current data set, in order to verify and refine our conclusions. One starting point would be to examine the distinct populations identified through the population genetic cluster analysis, with special emphasis on the Sambava locality, where C. medius and C. crossleyi are found in sympatry. Furthermore, a direct comparison of the speciation patterns found in dwarf lemurs and those known for mouse lemurs, with estimates of time divergence dates, would be valuable to answer questions about the mechanisms driving this species radiation in Madagascar.

Field samples from a total of 48 individuals across 14 localities in Madagascar were collected between March 2003 and May 2007 (Fig. 1, Table 1). Sampling sites were not evenly distributed across the island due to a bias in remaining forest cover 48. Sampling was concentrated on the larger fragments found in the eastern portion of the island and makes no claim to be exhaustive. A maximum of three individuals per site, amounting to 31 individuals, were sacrificed and preserved as morphotypes and are now housed at the Département de Biologie Animale de l'Université d'Antananarivo, Madagascar. Tissue samples from internal organs (liver, kidney and spleen) and muscle tissue were stored in 70% EtOH. An additional 17 individuals were caught using Sherman live traps. One hundred traps were set along two to three transects for an average of 11 nights per site and baited with banana. Tissue for molecular analyses of these individuals was obtained by ear clipping after animals were anesthetized with GM2 49. External morphological measurements were taken from all 48 individuals, while internal measurements were only available for the 31 morphotypes. Animals were released at the site of capture at dusk on the following day. A total of 44 additional tissue samples were collected from specimens in three European museums: the Muséum National d'Histoire Naturelle, Paris (MNHN); Naturalis – Nationaal Natuurhistorisch Museum, Leiden, and the Natural History Museum, London (NHM). These specimens are the same individuals studies in the taxonomic revisions of Groves 24. Small amounts of dried tissue were taken from skulls and in a few cases from skins. Of the 44 museum samples taken, we were able to include 17 in the final mtDNA analyses (Table 2). The presumed sampling sites of museum samples are marked with triangles on the map in Fig. 1. A total of 24 published Cheirogaleus haplotypes were incorporated into the analyses of the mtDNA data set (Table 3). Sequences from Mirza zaza, Microcebus berthae, M. murinus, and M. ravelobensis, which serve as representatives of other major cheirogaleid lineages 50, were used as outgroups to root the phylogenetic trees. All previously published GenBank sequences are listed in Table 3 and the sampling localities for Cheirogaleus sequences are marked with squares on the map in Fig. 1. All sequences generated for this study were deposited in GenBank under accession numbers EU825210-EU825610.

Genomic DNA was extracted from both field and museum tissue samples using the QIAamp™ DNA Mini Kit for DNA purification (QIAGEN) following standard protocol. For the field-collected samples two mitochondrial [cytochrome b (cytb), cytochrome oxidase II (cox2)] and three nuclear loci [adenosine receptor A3 exon 2 (adora3), alpha fibrinogen intron 4 (fiba), and von Willebrand Factor intron 11 (vWF)] were amplified, using primers and annealing temperatures given in Table 5. Amplifications were either carried out in 10 μl reactions containing a final concentration of 0.25 μM of each primer, 3 mM MgCl₂, 0.25 mM dNTPs, 1× amplification buffer and 0.025 U/μl taq (Jumpstart, Sigma) or in 30 μl reactions containing a final concentration of 0.33 μM of each primer, 3 mM MgCl2, 0.166 mM dNTPs, 1× amplification buffer and 0.033 U/μl taq (Biotherm, Genecraft). Cycling conditions were 35 cycles of 30 s denaturation, 45 s annealing and 1 min. elongation except for two loci. The adora3 fragment only needed 45 s elongation due to its short fragment length and the two overlapping cytb fragments were amplified using 1 min. for all three steps. Amplification of the museum samples was carried out in 30 μl reactions containing a final concentration of 0.33 μM of each primer, 2 mM MgCl₂, 0.166 mM dNTPs, 1× amplification buffer and 0.033 U/μl taq (Biotherm, Genecraft). Cycling conditions were 50 cycles of 1 min. denaturation, 1 min. annealing and 30 s elongation. All amplifications of museum samples were verified by replication in a second independent lab. Wax-mediated Hot Start PCR was used on all 30 μl reactions to increase specificity and yield. PCR products were purified employing Montage™ PCR Centrifugal Filter Devices (Millipore) according to manufacturer's instructions. For a little more than half of the nuclear sequences, due to multiple polymorphic sites in heterozygous individuals, PCR products were cloned into a pGEM vector (pGEM™-T EasyVector System I, Promega), averaging three clones per polymorphic sequence. Both strands of all PCR products were sequenced, using the respective primer pair used for amplification and standard vector primers M13F and M13R for the cloned products, employing the BigDye™ Terminator v1.1 Cycle Sequencing Kit (Applied Biosystems) on an ABI Prism™ 3100-Avant-Genetic Analyzer (Applied Biosystems).

Sequences were edited and aligned using Sequencher 4.7 (Gene Codes Corporation, Ann Arbor, MI, USA) and manually checked by eye. Subsequently, sequences were collapsed into haplotypes using MacClade 4.05 51. Haplotype data sets were used for all subsequent Maximum Likelihood (ML) and Bayesian phylogenetic analyses. For all mtDNA analyses cytb and cox2 fragments were concatenated, resulting in an alignment of 1824 bp. GenBank, museum and field sample sequences were collapsed into haplotypes separately, since, due to missing data, unambiguous assignment was not possible. Uncorrected "p" distances for the cytb locus were calculated as implemented in PAUP* v4.0b10 52. Optimal nucleotide substitution models for each locus were chosen using the Akaike Information Criterion (AIC) as implemented in Modeltest v3.7 53. All ML analyses were conducted using a genetic algorithm approach in Garli v0.951 54. In Garli only the model specifications settings were adjusted according to the respective data set; all other settings were left at their default value. Ten replicates were run for each data set to verify consistency in log likelihood (lnL) scores and tree topologies. Maximum-likelihood bootstrap percentages (BP) were estimated in Garli by performing 500 pseudoreplicate runs on each nuclear data set and 100 pseudoreplicates on the mtDNA data set. PAUP* v4.0b10 was then used to calculate a majority-rule consensus tree for each data set.

Bayesian analyses were conducted on a concatenated cytb+cox2 mtDNA data set and on the individual nuclear loci using MrBayes v3.1.2 5556. For the mtDNA analyses, a partitioned analysis was performed treating the cytb and cox2 genes as separate partitions, each with their own DNA substitution models. In all analyses we used four Monte Carlo markov chains (MCMC) with the default temperature of 0.2. Analyses were run for ten million generations with tree and parameter sampling occurring every 100 generations. Flat priors were assumed for the model parameters including the proportion of invariable sites and the gamma shape parameter of rate variation among sites. The first 25% of samples were discarded as burnin, leaving 75,001 trees per run. The adequacy of this burnin and convergence of all parameters were assessed by examining the uncorrected potential scale reduction factor (PSRF) 57 as calculated by MrBayes v3.1.2 5556, which should approach 1 as runs converge and visual inspection of the trace of the parameters across generations using the software Tracer v1.3 58. Posterior probabilities (PP) for each split and a phylogram with mean branch lengths were calculated from the posterior density of trees using MrBayes v3.1.2 5556. Phylogenetic trees were visualized using TreeEdit v1.0a10 59 and FigTree v1.0 60.

Statistical parsimony haplotype networks were constructed for each nuclear locus using the program TCS version 1.21 61 for each individual nuclear locus. A 95% connection limit was used and gaps were treated as missing data.

A Bayesian population assignment test implemented in Structure v2.2 62 was used to infer population structure based on a combined genotypic matrix from all four loci (adora3, fiba, vWF, and mtDNA) and also on the three locus nuclear matrix. An admixture model was used with correlated allele frequencies and no linkage among loci. For each number of populations assumed (K = 1 to K = 10) we performed 50 replicate runs. The four-locus data set was run with a burnin of 500,000 MCMC generations and 4 million subsequent generations. The three locus nuclear data set was run with a burnin of 250,000 generations and 1 million subsequent generations. The adequacy of the burnin and subsequent length of the MCMC chain was checked visually by plotting the parameters α and the lnL against the number of generations. In order to detect the favored number of genetic groups, an ad-hoc statistic ΔK 40 was calculated. In addition, a pairwise comparison of the 50 runs for each K was carried out using the perl script Simcoeff 63. This procedure is based on the estimated membership fractions generated by Structure for a given K. The similarity coefficient for a pair of structure runs reflects the proportion of identical membership of individuals assigned through the Monte Carlo process. The proportion of runs resulting in 95% of the coefficients being equal is used to assess the stability of the cluster allocations.

The clustering pattern for the run with the highest probability (estimated log probability of the data) for K = 2 to K = 6 was visualized using Microsoft^® Excel^® v11.3.3. Membership coefficients with posterior probabilities under 0.05 were disregarded and proportionally added to the remaining membership coefficients. Therefore, in Fig. 6, some individuals show membership in fewer than K clusters.