The overall phylogenetic relationships as deduced from the concatenated dataset are consistent with the now broadly accepted branching pattern of the mammalian tree 1 (Figure 1). The superordinal clades Euarchontoglires, Laurasiatheria and Afrotheria are highly supported, and within these clades most bootstrap percentages and posterior probabilities are also high. Afrotheria is now generally accepted as a natural group since molecular studies unanimously support its monophyly, using various methods 14272829. In contrast, until now only few morphological synapomorphies, notably placental morphology 30, an increase in number of thoracolumbar vertebrae 31, and testicondy 32, appear to support this grouping. Afrotheria are divided into Paenungulata on one hand and the three remaining afrotherian orders (Afrosoricida, Macroscelidea and Tubulidentata) on the other hand. The most probable hypothesis concerning these remaining orders is their grouping within a clade called Afroinsectiphillia 1328 within which the internal relationships remain unclear.

Within the paenungulate clade the Tethytheria (elephants + sea cows) are strongly supported by morphological and complete mitochondrial genome data 3334. Nuclear genes are ambiguous about this relationship and left the phylogenetic affinities between the three paenungulate orders essentially unresolved 1143536. Our concatenated tree shows for the first time, based on nuclear genes, strong support for one of the three possible hypotheses: the grouping of Hyracoidea with Sirenia (PP = 0.99 and BP = 89). Bootstrap trees supporting alternative hypotheses exclusively group elephant with hyrax (BP = 11); Tethytheria is never recovered. All four genes independently support this result; the high support for the sea cow + hyrax grouping is therefore expectedly due to the synergy of these non-conflicting informations. To test whether our extensive taxon sampling within Tenrecidae may have improved the phylogenetic accuracy 3738, all tenrecs but one (Tenrec ecaudatus) were removed from a new analysis. The results did not differ much; support for the Sirenia/Hyracoidea clade dropped negligibly in the concatenated analyses (PP = 0.98 and BP = 86). Interestingly, in a retroposon insertion analysis, Nishihara et al. 4 found one insertion supporting exclusively the grouping of hyrax with dugong. These authors dismissed the apparent synapomorphous hyrax-sea cow insertion as homoplastic, in favor of the morphological evidence for Tethytheria.

Similarly, the relations between the afroinsectiphillian orders have not yet been clarified, and conclusions vary in different studies. Mitochondrial data give highly inconsistent results 3439, while mixed data tend to group golden moles and tenrecs with elephant shrews, together being the sister group of aardvark, with rather strong support 13540. Our data also support these results, as the Afrosoricida/Macroscelidea clade (= Afroinsectivora) is displayed with high confidence (PP = 1.00 and BP = 93), and Tubulidentata is found to be the sister group of this clade (PP = 1.00 and BP = 95). With a smaller dataset (only one tenrec) the support for the Afrosoricida/Macroscelidea clade slightly increased (PP = 1.00 and BP = 96). Hence, enlarged taxon sampling cannot explain our strong phylogenetic results within the afrotherian clade. All four genes separately displayed Afroinsectiphillia either as paraphyletic or weakly supported therefore the present results are not due to gene sampling biases. The retroposon analyses of Nishihara et al. 4 proposed the grouping of golden moles, tenrecs and aardvark, to the exclusion of elephant shrews, on the basis of two shared retrotransposons.

The large-eared tenrec (G. aurita) has been included until now in only two molecular studies, by Olson and Goodman 17 and by Asher and Hofreiter 18. These two studies found two different results concerning its phylogenetic position. The first study, comprising three mitochondrial genes (ND2, 12s rRNA and tRNAvaline) and one nuclear marker (vWF exon 28), displayed, in a parsimony framework, the large-eared tenrec as the most basal of all Malagasy tenrecs. This result was not influenced by the inclusion of morphological characters in the analyses. Asher and Hofreiter 18, using exon 10 of the GHR gene and morphological data, found Geogale nested within the Oryzorictinae, as sister group of the Microgale/Limnogale clade.

In the present study we also sequenced GHR exon 10 and vWF exon 28, and in addition the intronless gene for ADRA2B and the first exon of AR. For all genes separately the results were congruent in placing Geogale as sister group of the Oryzorictinae (Figure 2), although not always strongly supported: ADRA2B: PP = 1.00, BP = 96; AR: PP = 0.77, BP = 86; GHR: PP = 0.64, BP = 59; vWF: PP = 0.93, BP = 61. Concatenation of the four genes led to a stronger support for this node: PP = 1.00 and BP = 93 (Figure 1). The position of G. aurita as sister group of the Oryzorictinae (Oryzorictes, Limnogale, Microgale) seems thus strongly supported. However, the KH- and SH-tests (Table 1) did not completely confirm the strength of our results, showing that placing Geogale as the most basal Malagasy taxon (Olson and Goodman's hypothesis) was indeed significantly worse than our best tree, but placing it within the Oryzorictinae (Asher and Hofreiter's hypothesis) did not significantly change the likelihood of the topology.

The differences with the results of Olson and Goodman 17 probably stem from the fact that we did not use the same phylogenetic methods and datasets, even though one of our markers was in common (vWF exon 28). However, their vWF (exon 28) sequences are not yet available in public sources like GenBank to be compared with ours. The different position of Geogale in the tree of Asher and Hofreiter 18 is more difficult to explain. Remarkably, their Geogale GHR sequence (Acc. Nr.: DQ202287) displays 18 differences with ours (10 synonymous and 8 non-synonymous substitutions). No mutations leading to unusual amino acid changes that might indicate sequencing errors could be detected. To try and explain the different Geogale GHR sequences we calculated Ka and Ks for each sequence pair of Malagasy tenrecs. The results showed that the sequence divergence between the two Geogale specimens was greater than between some of the other tenrec genera, like Echinops/Setifer and Limnogale/Microgale (Table 2). Moreover, the new Geogale sequence from this study was slightly more divergent in most comparisons than the one from the database (Table 2). This genetic diversity within Geogale could reflect that this genus might contain in fact more than one species. It may also be mentioned that the museum specimen used by Asher and Hofreiter 18 was collected at the southwest coast of the island (Lamboharana, voucher number MCZ 45044), whereas our specimen (voucher number MVZ mammal # 220648) was sampled in the central west in the Menabe area. Considering photos of living Geogale available to us from the south-west (by W. R. Branch) and the central-west (by R. Nincheri and ourselves), the central western specimens appear to have a less golden-colored fur and in general a more gracile habitus, but it is unclear whether this may reflect a difference between coloration of adults versus subadults. Clearly, a detailed taxonomic study is needed to confirm whether these differences are constant and the populations may represent two distinct species. Furthermore, a single record of Geogale exists also from the east coast near Fenoarivo. This specimen has been described as subspecies Geogale aurita orientalis by Grandidier and Petit 41, but the status of this taxon has remained obscure. It may be a candidate nomen to be elevated to species rank if Geogale aurita is demonstrated to consist of more than one species.

Further phylogenetic analyses of the GHR dataset, including both Geogale sequences or removing all segregating sites between the two sequences, led to the same result as obtained by Asher and Hofreiter 18, i.e. Geogale nested within the Oryzorictinae. The phylogenetic position of Geogale as sister group of Oryzorictinae was only obtained when our sequence alone was used. However, both Geogale sequences always grouped together, confirming the identity of our sequence. These results, in combination with the fact that the Oryzorictinae/Geogalinae clade radiated very fast, might make it difficult to reach a final consensus on the evolution of Geogale.

From a morphological point of view the phylogenetic relation between Geogale and the Oryzorictinae has never been clear. Although most studies gave unresolved results [10, Olson [1999] in 1718], two were concordant with ours 4243, while none has ever argued that Geogale was either the sister group of all Malagasy tenrecs or the sister group of the Limnogale/Microgale clade. Salton and Szalay 43 reached the conclusion that the tarsal morphology of Geogale warrants its status as a separate subfamily, and suggested its closer affiliation with Oryzorictinae than with Tenrecinae.

Three genera of fossil tenrecids – Erythrozootes, Protenrec and Parageogale – from the Kenyan and Namibian Miocene (16–24 Mya; Million years ago) have been discovered until now 444546. As Parageogale is thought to be the sister group of the extant Geogale aurita 45, these data would suggest a more complex dispersal history than the "one time dispersal event" deduced from the monophyly of Malagasy tenrecs. Asher and Hofreiter 18 were the first to include these three fossil tenrecids in a phylogenetic framework. Their result confirmed the position of the Kenyan fossils as Geogale's closest relatives. However, alternative hypothesis (e.g., monophyly of the Malagasy tenrecs) could not be ruled out indicating the uncertainty of the Parageogale/Geogale affinity. Recent studies have argued that the sweepstakes dispersal model (dispersal with small and random probability of success) from Africa to Madagascar suffers from many inconveniences, among which the fact that prevailing winds and currents between Africa and Madagascar would be much more likely to favor transports from the island to the African continent, rather than the reverse route 4748. Therefore, if a second dispersal event ever occurred it was most probably from Madagascar to Africa. Olson and Goodman 17 suggested a basal position of Geogale among Malagasy tenrecs and argued that, if true, this would only imply a minimum of two dispersal events, whereas any other scenario would require at least three. However, a back dispersal of Parageogale from Madagascar to Africa would only assume a second dispersal event, independent of the phylogenetic position of Geogale.

Due to its semi-aquatic life style, shared with the African Potamogalinae, the determination of the phylogenetic relationship of Limnogale, the web-footed tenrec, has led to controversies. Its specialized morphological features brought some authors to the conclusion that Limnogale was either sister group of the Potamogalinae 10 or sister group of all other Malagasy tenrecs 42, the semi-aquatic behavior then being seen as an ancestral state and a key element to facilitate over-water dispersal. In contrast, other morphological studies challenged this view by affirming that Limnogale had closer relationships to the shrew tenrecs (Microgale), and that the semi-aquatic behavior was an example of convergence acquired twice during tenrec evolution [Guth et al.[1959] in 17, Olson [1999] in 17]. This strong affinity between Limnogale and Microgale has recently also been supported by a study of hind limb muscles 49. These authors argue that Limnogale may have been derived from a Microgale-like terrestrial ancestor. Molecular studies have now confirmed this last hypothesis 161718. Supporting the hypothesis of Olson and Goodman 17, our study shows that the semi-aquatic Limnogale is actually nested within the shrew tenrec genus and not a sister clade of it (Figure 1), now with more elaborate analyses and strong support from four nuclear genes.

The phylogenetic supports displayed in the present study are quite low, even with the concatenated dataset (PP = 0.67 and BP = 59), probably due to the fact that the Microgale/Limnogale clade may have radiated very fast (Figure 1). Only one gene, GHR, presents a high PP of 0.99 for the cluster of Microgale cf. parvula/Limnogale mergulus (Figure 2). The sequencing of more shrew tenrec species (a total of 21 species has been recorded 5678) might help to resolve this issue, and subsequently to understand the morphological evolution of the aquatic specialization of the web-footed tenrec.

Only three studies have previously assessed the timing of tenrec diversification, mainly to understand their colonization pattern 131416; none comprised a taxon sampling broad enough to delineate the successive tenrec speciation events. The study by Douady et al. 13 was based on a linearized tree method and suggests an early diversification of Tenrecs as compared to the other studies (for the present study see Figure 3), which are based on Bayesian methods and partially overlapping gene sampling (Table 3). Consequently, the results of the latter three studies are, as can be expected, rather similar. The present study, with the broadest taxon and gene sampling, estimates the tenrecs/golden mole split at 69 ± 4 Mya, followed by the divergence between African and Malagasy tenrecs at 47 ± 4 Mya. The Malagasy tenrec radiation began 29 ± 3 Mya, and several diversification events spread over time gave rise to the totality of Malagasy tenrec genera around between 20 ± 1 Mya and 7 ± 1 Mya (Table 3 and Figure 3). These datings are slightly older than previously calculated. The only gene difference between this study and Poux et al. 16 is the inclusion of the GHR gene. Removing it from the calculations led to dates even a little older and with wider confidence intervals (Table 3).

Because the GHR influence on the dating was very small, the difference in taxon sampling between the two studies might be responsible for the different outcomes 50. In the present study carnivores and primates were less extensively sampled, whereas Afrosoricida were better represented than in Poux et al. 16. We therefore compared for these three clades the age inferences in Poux et al. 16 and in the present study, with or without GHR (Table 4). The conclusion is that the age of a given node tends to become older when the taxon sampling around this node (or descending from it) increases. This phenomenon has already been described by Yoder and Yang 51 when assessing the timing of evolution of mouse lemurs. They suspected that these incongruences were due to the model used 52, which breaks down the path from a tip of the tree to the root (or ancestral node) into identically distributed segments. Such a prior would tend to push divergence time within the clade under study towards unrealistically old ages. Comparing the priors of divergence times between both large and small datasets, they reached the conclusion that the too old priors of the larger dataset had influenced the posterior estimates, which became older as well. This also is the pattern we can see comparing the priors of Poux et al. 16 with the ones of the present study (dataset without GHR). In both studies the time estimate differences were not dramatic, but they could have a problematic effect for studies requiring more precise estimates.

The influence of data partitioning was tested as well. The ages of the nodes in the phylogenetic tree increased with the number of partitions (Figure 4A), and the smallest standard deviations (and therefore confidence intervals) were reached for the less extreme numbers of partitions (Figure 4B). However, for the present study, differences in taxon sampling or partitioning did not affect our conclusions, as the various analyses displayed fairly similar results, showing reciprocal overlaps. This is to our knowledge the first time that the influence of data partitioning on dating results has been empirically pointed out. More investigations are needed to generalize and clearly understand the underlying causes of this result. One might however suppose that the differences between the various partitions could increase with the number of genes included in an analysis. Consequently, these results show that it is important, in order to calculate datings as accurately as possible, to select the right manner of partitioning the data: too few or too many partitions might lead to biased results.

To exclude the possibility that individual calibration constraints may bias our dating analyses, we repeated them after removing each calibration point in turn following 16. Hereby we could check whether the excluded calibration constraint was accurately estimated by the remaining ones. All datings remained highly congruent when any of the six calibration points was removed. The average percentage difference between the main analysis and the ones with only 5 constrained nodes ranges between 0.1 and 0.8 percent. Only the paenungulate calibration seems to have a somewhat larger impact on the dating as its removal from the analysis increases the estimated node age by 4.8 percent. This influence is however too slight to have an impact on our conclusions (Additional files 1 and 2). Moreover, the calibrations were reciprocally compatible: the remaining five calibrations always recovered a posterior estimate (± SD) for the excluded node within the time window independently obtained from the corresponding fossil evidence (Additional files 1 and 2).

Since Geogale has been hypothesized by Olson and Goodman 17 to be the first Malagasy tenrec genus to have diverged, its absence from Poux et al. 16 was a problem for drawing final conclusions about tenrec colonization timing. It now appears that Geogale is nested within the Malagasy tenrec clade, and therefore plays no role when estimating the period of colonization. Consequently, the window of colonization of Madagascar by tenrecs could not be narrowed. As previously concluded in Poux et al. 16, the tenrec colonization time completely overlaps with the hypothetical time of existence of a land bridge crossing the Mozambican channel (26–45 Mya; 53) (Figure 3), which however is highly controversial 54.

Adaptive radiation often occurs when a species is introduced into a new environment, such as an island. One might therefore expect that the majority of the diversification events within the Malagasy tenrecs would have occurred soon after colonization. However, no such pattern of a diversification burst can be seen at the root of the Malagasy tenrecs, and speciation events seem to be spread through time (Figure 3). This could result from two possible scenarios: either Malagasy tenrecs may actually have experienced a fast adaptive radiation, but most of the resulting taxa are now extinct, or all genera appeared indeed at different periods as a result of a slower speciation rate than expected in case of adaptive radiations. Morphologically, one might speak about adaptive radiation of Malagasy tenrecs, but these morphological adaptations do not seem to have developed within a short time span just after the colonization of Madagascar. The most striking example is the semi-aquatic specialization of the genus Limnogale, which dates at most from 11 Mya, i.e., 20–38 My after the colonization of the island. The genus Microgale is by far the most speciose amongst tenrecs, being represented by 21 species 5678, while the remaining genera may not comprise more than one species. The acceleration of the molecular evolutionary rates on the internal branches leading to and within this genus (calculated with MULTIDIVTIME on the tree presented in Figure 1), associated with both a poor phylogenetic resolution between the few Microgale species (Figs 1 and 2) and its recency among the tenrec genera (Figure 3 and Table 5), suggests that there has been a fast radiation around 11 Mya that gave rise to the current diversity of Microgale. It is interesting to note that the other two endemic mammalian Malagasy genera for which radiation times have been assessed apparently diverged around the same time as Microgale: Eulemur at 9.7 Mya and Microcebus at 8.7–12 Mya 51. However, not enough data are yet available to confirm this parallel radiation phenomenon.

Even though the colonization of Madagascar by tenrecs might have taken place during the Eocene, the radiation of the extant species started after Madagascar reached its current geographical subtropical location during the early Oligocene 55, with warmer climatological conditions probably similar to the actual ones 56. The colonization of Madagascar by carnivores and rodents took place at the end or just after the Oligocene, around 20–23.5 Mya for rodents, and 19–26 Mya for carnivores (data taken from 16 in order to compare results inferred from similar datasets and methods). These dates are quite close to the periods of appearance of extant tenrec genera: the radiation of Tenrecinae and the split between Tenrec and Hemicentetes occurred 20 ± 2 Mya and 16 ± 2 Mya, respectively; Geogale split from the Oryzorictinae 24 ± 3 Mya; and Oryzorictes separated from Microgale 22 ± 3 Mya. So five out of the seven tenrec genera (Limnogale is taken here as a Microgale) diverged soon after the colonization of Madagascar by carnivores and rodents. These new colonizations may have altered the ecological conditions, and thereby induced speciation within tenrecs, either by predation pressure (carnivores) or by interspecific niche competition (rodents).

Fragments of the intronless gene of the alpha 2B adrenergic receptor (ADRA2B), of exon 1 of the androgen receptor (AR) gene, of exon 10 of the growth hormone receptor (GHR) gene, and of exon 28 of the von Willebrand factor (vWF) gene were amplified and sequenced. These genes were selected because (i) they are located in the nuclear genome, as single-copy genes (in at least human and mouse), (ii) a considerable number of sequences are already available for all four genes and have been useful in mammalian phylogeny, and (iii) they are functionally and genetically unrelated. We selected for each of the four genes 38 mammalian species to represent (i) all genera of Malagasy tenrecs, and at least two species of the very diverse genus Microgale, in order to assess the phylogenetic position of Limnogale, (ii) the continental African sister group (Potamogalinae) of the Malagasy tenrecs, (iii) groups needed for multiple calibrations of the molecular clock, (iv) at least one species from each eutherian order (but for Pholidota), and (v) appropriate marsupial outgroups. A total of 19 new sequences were obtained, and complemented with 134 sequences from GenBank (Table 5).

Genomic DNA was isolated from ethanol-preserved tissue, following the protocols of the Wizard^® SV Genomic DNA Purification System (Promega). Fragments of the ADRA2B and AR genes were amplified using previously published primers 1657. New primers were designed for vWF and GHR (see Additional file 3). For these last genes PCR reactions were performed on 50–200 ng DNA with Expand DNA polymerase (Expand High Fidelity PCR system, Roche) using the following program: 2 min at 94°C; 30–35 cycles of 15 sec at 94°C, 1 min at 60°C and 1 min 30 sec at 72°C; and a final step of 2–10 min at 72°C. DMSO (1.3 – 2.5%) and/or betaine (1 M) was added for some samples. PCR products were purified from a 1% agarose gel, using GFX™ PCR DNA & Gel Band Purification Kit (GE Healthcare), and reamplified if necessary. Gel-extracted PCR products were sequenced directly on a 3730 96-capillary sequencer (Applied Biosystems). Internal primers were used to get complete sequences of both strands.

Sequences were assembled and aligned with the ED editor of the MUST package 58, and manually adjusted taking amino acid properties in consideration. Amino acid repeats and sites not sequenced or gapped in more than 25% of the taxa were excluded from analysis. This resulted in a dataset of 1,101 bp for ADRA2B, 1,161 bp for AR, 852 bp for GHR, and 1,173 bp for vWF. The full data matrix is available from Treebase (accession number: M3679). Phylogenetic reconstructions on each gene separately and on the concatenated dataset were performed by maximum likelihood (ML) with PAUP*, version 4b10 59, and by Bayesian analyses with MRBAYES, version 3.1.2 60. The best fitting model under the ML criterion was selected from the "Akaike Criterion" output of MODELTEST, version 3.7 61. The ML analysis was conducted using a loop approach to estimate the best tree and the optimal likelihood parameters. With this approach parameters and best tree are re-estimated until they reach stability. Node stability was estimated by 100 non-parametric bootstrap replicates 62. A major advantage of Bayesian phylogenetic inference is the possibility of partitioning the data, giving each partition its own best fitting model of sequence evolution. However, overpartitioning may introduce unnecessary sampling variances which could influence the phylogenetic estimates. For the twelve possible codon partitions (each codon position of each gene) MODELTEST was used to calculate the best fitting model of sequence evolution. As further explained in Table 6, codon partitions with similar models and model parameters were merged, resulting in nine partitions for the Bayesian analyses. Two runs of four Markov chains were calculated simultaneously for 1,000,000 generations with initial equal probabilities for all trees and starting with a random tree. Tree sampling frequency was each 20 generations, and the consensus tree with posterior probabilities was calculated after removal of the first 25% of the total number of trees generated, corresponding to 12,500 trees. The average standard deviation of split frequencies between the two independent runs was lower than 0.01.

To assess the stability of the phylogenetic position of Geogale aurita, our result was compared, according to both Kishino and Hasegawa 63 and Shimodaira and Hasegawa 64 (using RELL bootstrap as well as full optimization methods), to the hypotheses of Olson and Goodman 17 and Asher and Hofreiter 18. Furthermore, Ka (i.e., number of nonsynonymous substitutions per nonsynonymous site) and Ks (i.e., number of synonymous substitutions per synonymous site) of pairwise tenrec sequences were calculated using the program CODEML from the PAML package 65 in order to assess the molecular divergence between the two Geogale GHR sequences and compare it with the level of molecular divergence displayed within the Malagasy tenrec clade.

We used the Bayesian approach 66 as implemented in the MULTIDIVTIME program package 52, which relaxes the molecular clock by allowing continuous autocorrelation of substitution rates among the branches of the phylogenetic tree. The concatenated sequence dataset was partitioned into the same nine categories as for the Bayesian phylogenetic analyses, and branch lengths were calculated under the F84 + Γ model of sequence evolution, which is the most complex model available in MULTIDIVTIME. Each of the described analyses was run twice in order to assess the consistency of the results. The prior for the root was set at 100 Mya, however, analyses with 65 Mya, 80 Mya and 120 Mya as prior age were also performed in order to estimate the impact of the root prior on our results. For each node, we calculated the variance of the estimated ages over all the runs. A maximal variance of 2*10^-4 was found showing that changing the root prior does not influence age estimates. Markov Chain Monte Carlo analyses were run for 1,000,000 generations after a "burn in" of 100,000 generations. The chains were sampled every 100 generations. To assess the influence of a particular partitioning on the dating results, we performed additional analyses using four partitioning schemes: without partitioning, with nine partitions following the results of MODELTEST, with five partitions following the results of ESTBRANCHES using the F84 + Γ model, and with a maximum number of partitions (i.e., twelve). The results of these analyses were close to each other. Notably, all datings for the nodes of interest remained within the 95% credibility intervals of the datings obtained in the analysis using five partitions.

Six well established fossil constraints on divergence times were used: (i) a minimum of 54 and a maximum of 65 Mya for the base of Paenungulata 67; (ii) a minimum of 50 and a maximum of 63 Mya for the split between feliform and caniform Carnivora 4568; (iii) a minimum of 54 and a maximum of 58 Mya for the split between hippomorph and ceratomorph Perissodactyla 69; (iv) a minimum of 55 and a maximum of 65 Mya for the base of Cetartiodactyla 70; (v) a minimum of 37 Mya for the split between ochotonids and leporids 45; (vi) a minimum of 60.5 and a maximum of 100.5 Mya for the divergence time between rodents and primates 71. To assess the reciprocal consistency of all calibration points we used the cross-validation method described in 16. In this method each calibration point is removed in turn and the remaining calibration points are used to estimate its age. Calibration points, for which the estimated and paleontological dates are not congruent, are considered as inconsistent and are consequently removed from the analyses.