Phylogenetic analyses using maximum parsimony (MP), two different methods of maximum likelihood (ML) and Bayesian inference (BI) all recovered the same hypothesis for intergeneric relationships, which resolves Mustelidae into seven primary divisions that include four major clades and three monotypic lineages (Figure 1). Otters (Aonyx, Enhydra, Hydrictis, Lontra, Lutra, Lutrogale and Pteronura) form a clade (node 7) that is sister to a clade comprising mink and true weasels (Mustela and Neovison; node 16). These clades, in turn, are sister to a clade that includes weasel-like species with aposematically colored pelage (Galictis, Ictonyx, Poecilogale and Vormela; node 27). Next, ferret-badgers (Melogale) are a monotypic lineage (node 32) that is sister to these three combined clades. The fifth major clade (node 33) comprises two subclades, one containing hog-nosed and Eurasian badgers (Arctonyx and Meles) and the other containing tayra, wolverine and martens (Eira, Gulo and Martes). Finally, as the earliest divergences in the tree, the American badger (Taxidea) and honey badger (Mellivora) form successive monotypic lineages sister to all other mustelid genera. Intergeneric relationships observed are largely congruent with those recovered in recent analyses of mitochondrial and nuclear DNA sequences 10111314151617. In contrast, our topology is highly incongruent with a cladistic analysis based on morphology 8. For example, Bryant et al 8 found that Melogale was reconstructed as sister to all remaining mustelids and that Eira, Gulo and Martes were polyphyletic. Except for the monophyly of otters, skunks (including Mydaus), and Arctonyx and Meles, all other nodes in the Bryant et al 8 tree had low bootstrap support values (<50%). Furthermore, alternative topologies were recovered in the Bryant et al study when certain characters were weighted differentially.

Regarding relationships within the four major clades, otters are resolved into three primary lineages, whose relationships are congruent with previous analyses based on fewer DNA sequence data 1213: one containing Old World river otters and the sea otter (Aonyx, Lutrogale, Lutra, Enhydra and Hydrictis; node 9), a second containing New World river otters (Lontra; node 14) and a third containing the monotypic giant otter (Pteronura). Relationships of Enhydra and Hydrictis were unresolved in the Bayesian consensus tree (Figure 1), but were resolved differently in MP and ML analyses. With MP, Enhydra and Hydrictis were resolved as sister taxa, whereas with ML, they were resolved as successive sister lineages to remaining Old World otters, with Hydrictis forming the first basal split. However, neither of these relationships was well supported (<50% MP and ML bootstrap). Two clades comprise the true weasel and mink clade (Mustela and Neovison), with the New World American mink (N. vison) and long-tailed weasel (M. frenata) in one clade (node 21) and all other sampled Mustela species in the second clade (node 17). Within the latter clade, M. nudipes and M. strigidorsa are sister to a clade comprising species largely distributed in temperate regions of the northern hemisphere. Despite the similar ecology, American mink (N. vison) and European mink (M. lutreola) are distantly related, as found in previous studies 143334. Species of true weasels and mink have been divided into five 34 or nine subgenera 35 based on morphological criteria. Given the taxa we have sampled, our phylogeny suggests that only one proposed subgenus constitutes a natural group, Putorius, containing the steppe polecat (M. eversmanni), European polecat (M. putorius) and the North American black-footed ferret (M. nigripes; node 25). Further, Mustela is paraphyletic with respect to Neovison and suggests that the placement of the American mink in the separate genus Neovison may not be warranted despite the observed differences in karyotype and morphology between this taxon and other species of Mustela35. In the third major clade, grisons of Central and South America (Galictis; node 31) are sister to a clade containing the marbled polecat (Vormela) of Asia and three African species, the Libyan striped weasel (Ictonyx libyca), zorilla (I. striatus) and striped weasel (Poecilogale albinucha; node 28). Monophyly of these genera is of interest because all exhibit an aposematically colored pelage combined with defense behaviors that include threat displays and excretion of pungent musk from enlarged anal glands 36373839. Placement of Vormela within this clade is congruent with a recent CYTb study 40. Interestingly, there is a north to south progression in branching order from Vormela to I. striatus and P. albinucha of sub-Saharan Africa, which renders Ictonyx paraphyletic. This suggests that P. albinucha should be placed into the genus Ictonyx Kaup, 1835, given the priority of the latter name over Poecilogale Thomas, 1883. As for the clade containing Eira, Gulo and Martes (node 34), the last genus is clearly paraphyletic, in agreement with other recent studies 11134142. Two New World species, the tayra (Eira) and fisher (Martes pennanti, subgenus Pekania) either comprise a clade or form successive lineages sister to a clade containing wolverine (Gulo) and the remaining species of Martes. Our phylogeny indicates that true martens (subgenus Martes) are monophyletic (node 38) and sister to yellow-throated marten (M. flavigula, subgenus Charronia). Within the subgenus Martes, MP and ML analyses resulted in different phylogenetic placements of M. americana and M. melampus relative to the (M. martes + M. zibellina) clade. In the MP tree, M. americana was sister to the clade (M. melampus (M. martes + M. zibellina)) whereas M. americana and M. melampus were joined as sister taxa in the ML tree. Relationships of M. americana, M. melampus and (M. martes + M. zibellina) were unresolved in the Bayesian tree, resulting in a trichotomy (Figure 1; node 39). Although more data is required to establish the precise branching order of the fisher, the strong support for this taxon being paraphyletic with the remaining species of Martes strongly suggests placement of the fisher in its own genus, Pekania.

Concatenation of 11,929 bp, with indels coded as missing, includes 2,959 (24.8%) variable characters and 1,876 (15.7%) parsimony-informative characters. When indels are coded for information content, the concatenation is reduced to 11,789 bp, with 3,045 (25.8%) variable and 1,917 (16.3%) parsimony-informative characters. The majority of internodes in our phylogeny have >90% bootstrap support (ML and MP) and 1.0 posterior probabilities (BI), see Figure 1 and Table 1, indicating that our phylogenetic hypothesis, given our data set, is robustly supported. Moreover, 14 internodes are supported by one or more synapomorphic indels (Table 1). Three clades, however, have low support in MP, ML and/or BI analyses: nodes 26, 33 and 35. Internodes associated with these clades have very short branch lengths (Figure 1), suggesting that these splits occurred rapidly. Alternatively, speciation may have occurred relatively recently, also resulting in inadequate phylogenetic signal, as among M. eversmanni, M. putorius and M. nigripes (node 26) as well as M. americana, M. melampus and (M. martes + M. zibellina). Many nuclear segments show no differences among species that comprise these groups, suggesting that there has been insufficient time for sequence differences to accumulate at these more slowly evolving loci. Indeed, the four species of Martes have been described as a superspecies complex of closely related, yet largely allopatrically distributed taxa 43. Regardless of the exact cause, short branch lengths of these internodes can result in a high number of anomalous gene trees (and gene tree discordance 44), leading to an incorrectly inferred species tree, especially when data from multiple data partitions is concatenated 45. Additional analyses, including population genetic-level sampling, will be needed to confidently resolve relationships among these recently evolved species, as in reference 33 for example.

Even so, internodes with low support may be stable or unstable, depending on the relative stability or instability of their constituent (terminal) taxa or leaves 46. Further, unstable taxa can influence nodal support in other parts of a phylogenetic tree 47. We measured leaf stability of all taxa using bootstrap trees from MP analyses and the program RadCon 48. For three different measures of leaf stability, Arctonyx collaris and Meles meles were the least stable taxa (see Additional file 1). We repeated MP and hill-climbing ML analyses (see the methods section) along with bootstrapping after excluding these two taxa, thereby reducing the data set to 44 taxa. Most bootstrap values in these analyses showed little or no change compared with the 46 taxa data set, but interestingly, support for two stem clades (nodes 3 and 4) increased from over 70% to more than 90%, as did average phylogenetic stability (Table 1). The lower stability of A. collaris and M. meles and their influence on bootstrap support at internodes 3, 4 and 33 may be caused by character conflict among different gene segments associated with these two taxa 49. However, we note that these nodes all have high or maximal posterior probabilities in BI analyses (Table 1).

With regards to subfamilial classification, our phylogeny clearly indicates that the Mustelinae, as both traditionally 5 and recently 7 conceived, is polyphyletic, as suggested by previous studies 81314. Our results also indicate that various genera of badgers (Arctonyx, Meles, Mellivora, Melogale and Taxidea), most of which are often placed in the Melinae 5, are also polyphyletic, supporting earlier conclusions based on morphology that these taxa are not closely related 85051. Instead, badgers with specific adaptations for fossoriality (Arctonyx, Meles, Mellivora and Taxidea) constitute a basal grade of lineages that are best recognized as distinct subfamilies. Overall, our phylogeny (Figure 1) is consistent with the subfamilial classification scheme recently proposed by Sato et al 15 and Fulton and Strobeck 11, which redefines the boundaries of traditional subfamilies (e.g. Mustelinae) as well as resurrects other subfamilies such as Helictidinae and Mellivorinae, thereby reflecting actual phyletic lines that have been difficult to resolve based on morphology alone.

Divergence times across the mustelid phylogeny were estimated using the uncorrelated relaxed lognormal molecular clock model calibrated simultaneously by eight fossil constraints. Use of this molecular clock model provides a measure of rate heterogeneity among lineages or how well data conform or deviate from a strict molecular clock 52. For the root age and crown age prior combination of 28.5 MYA and 24 MYA, the coefficient of variation (σ_r) averaged across three independent runs was 0.375 (95% highest posterior density (HPD): 0.274–0.486), which suggests that the data show rate heterogeneity among lineages (i.e. the concatenated data are evolving in a non-clocklike manner). We found greater rate heterogeneity among lineages when older root age and crown age prior combinations were used; for example, for 33.7 MYA and 28.5 MYA, averaged across three runs, σ_r = 0.504 (95% HPD: 0.381–0.631). Furthermore, a likelihood ratio test significantly rejects the ML phylogeny with an enforced molecular clock versus one without a molecular clock (-ln L_{with clock} = 55179.13 and -ln L_{without clock} = 55095.88; p < 0.001) thereby providing additional support that our data departs from clocklike behavior.

In general, mean posterior estimates as well as 95% HPD intervals of node ages were highly consistent among different combinations of root age and crown age prior combinations (Table 2). The root and node 1 (the divergence between Taxidea and remaining mustelids) were most sensitive to alterations in priors, because they showed the largest shift in divergence times (~5 million years) and little overlap in 95% HPD intervals in runs using older versus younger combinations of root age/crown age priors (Table 2). Nonetheless, sampling of the joint prior distribution by performing Markov chain Monte Carlo (MCMC) analyses without any data (using BEAST v1.4.2 52) suggested that the eight fossil calibration point priors (as well as the various root age and crown age combinations) did not have a strong influence on their estimated posterior distributions (see Additional file 2) and thus, estimated divergence times.

How do our estimates of divergence times compare with those based on previous molecular studies of mustelids 12141741? At nodes shared among the different studies, divergence times in previous studies are either a mix of younger, older and overlapping dates relative to those in our study 17 or they tend to be older in general 121441. For example, the estimated divergence time for the split between Neovison vison and the remaining taxa of Mustela was dated at 8.5–9.9 MYA, 6.6–9.5 MYA and 10–14 MYA in Sato et al 14, Marmi et al 17 and Hosoda et al 41, respectively. In our study, this split (node 16; Figure 2 and Table 2), which also includes M. frenata as sister to N. vison, is dated at ~6.0 MYA (95% HPD: 4.6–7.3 MYA). However, earlier studies differ in several important respects from our study. First, these studies were based on a smaller number of taxa and loci. For example, in Sato et al 14 and Hosoda et al 41, taxon sampling largely consisted of species from Martes and Mustela and was based on one and two loci, respectively. Undersampling of taxa and characters can bias divergence time estimates 535455 and reduces the power to detect rate variation among lineages 56. Second, previous studies all employed a single root age fossil constraint, treated as a hard bound, precluding an accurate estimation of error associated with this calibration point. In contrast, we employed multiple fossil constraints simultaneously and placed soft-bound priors on these constraints to account for the uncertainty associated with the fossil record, thereby making it possible to evaluate errors associated with divergence time estimates 5758. Third, the previous studies each estimated a substitution rate that was assumed to be constant across their respective phylogenies. However, our results indicate that substitution rates vary across different lineages of the mustelid phylogeny (see above), consistent with recent evidence from different groups that rates of substitution vary across lineages, even among closely related species 59.

Using the root age and crown age prior combination of 28.5 MYA and 24 MYA as a reference, our dating analyses indicate that, following the initial divergence of Taxidea in the Early Miocene (21.0 MYA, 95% HPD: 19.0–23.1 MYA), mustelids underwent two main bursts of diversification (Figure 2). The first burst occurred during a ~3.7 million year interval from the Middle to Late Miocene (12.5-8.8 MYA, 95% HPD: 13.9-7.5 MYA) and gave rise to most of the extant primary clades and lineages (nodes 2–6 and 33). The second and larger burst occurred during the Pliocene epoch (5.3-1.8 MYA) in which as many as 20 generic-level or specific-level lineages originated within a 3.5 million year span of time (Figure 2 and Table 2). Furthermore, results of our biogeographic analyses show that the majority of cladogenetic events associated with these bursts of diversification occurred in the Old World, largely in Eurasia (Figures 3 and 4; see below).

Paleoenvironmental and biotic changes driven by changes in climate during the latter half of the Neogene may have promoted the two bursts of cladogenesis within mustelids. Following the Mid-Miocene Climatic Optimum and onset of modern oceanic circulation (17-15 MYA), there is a marked cooling of the global climate near the end of the Middle Miocene that continues through to the Holocene 60. This period of cooling coincides with formation of a permanent Antarctic ice sheet in the Mid to Late Miocene and an Arctic ice sheet in the Pliocene 60. In addition, several major sea-level lowstands occurred during the Late Miocene and Pliocene, including the Serravallian sea-lowering event near the beginning of the Late Miocene, 11-10 MYA 6162. These changes in climate and sea level increased overall terrestrial aridity and seasonality, which in turn promoted a shift from closed vegetation habitats (tropical and subtropical forests) to more open vegetation habitats (woodlands and grasslands) 636465. By the early Late Miocene, plant and animal fossil evidence indicates that the Eurasian continent was a mosaic of vegetation types and generally more heterogeneous in vegetation structure relative to that in the Early to Middle Miocene 66. These changes in vegetation had a concomitant impact on faunal communities and may have fostered diversification in a variety of lineages, including mustelids, via geographic isolation, divergent selection among different habitats, and/or ecological opportunity through the creation of new niches or the reorganization of former niches. Interestingly, the initial radiation of the primary clades and lineages of mustelids coincides with a major faunal turnover in Western Europe (the middle Vallesian 'crisis') that affected many groups of mammals, including the Carnivora 6267. In fact, nearly half of carnivoran species that went extinct during this turnover were mustelids 68 and turnover in mustelids remained high throughout the Late Miocene in western Eurasia 69. Although evidence for faunal change in other parts of Eurasia is less clear, changes in habitat and extinction of earlier lineages of mustelids may have created ecological opportunities that fostered the initial burst of diversification of modern mustelids.

Further cooling and drying during the Pliocene, coincident with onset of high latitude glacial cycles 60, caused a dramatic expansion of low-biomass vegetation, including grasslands and steppe at midlatitudes and development of taiga at high latitudes of Eurasia and North America 636465. Coupled with these changes was diversification of prey species such as rodents (particularly muroid rodents) and passerine birds that exploited these new habitats, which in turn provided new niches for predators 6870. Part of the Pliocene burst of diversification of mustelids may have been promoted by this diversification of prey species, as species in genera such Martes (nodes 37–39 in Figure 2) and especially Mustela (nodes 17–23 in Figure 2) became specialized in hunting small prey such as rodents. Such a scenario is consistent with King's 71 hypothesis that evolution of small body size in Mustela was partly driven by adaptation to exploit abundant resources presented by rodent diversification during the Pliocene. We also note that diversification of four of the five species within the subgenus Martes (M. americana, M. martes, M. melampus and M. zibellina; node 39), which are all closely associated with taiga forest habitat 72, coincides with expansion of this type of habitat across the Holarctic during the Plio-Pleistocene 63. This finding is consistent with fossil evidence that indicates that taxa ancestral to these living species primarily evolved in forested habitats 4373. In Africa, the split between Ictonyx libyca and the clade containing I. striatus and Poecilogale albinucha (node 29) dated at 3.0–3.5 MYA (95% HPD: 1.8–4.6 MYA), marginally overlaps with a major increase in African aridity and climate variability that occurred 2.9-2.4 MYA, according to paleoclimatic and faunal evidence 7475. Moreover, this interval was accompanied by rapid radiation in several mammalian lineages such as bovids and hominids 7475. Our results suggest that divergence of I. libyca, which occurs in North Africa along margins of the Sahara desert, and the clade containing I. striatus and P. albinucha, which both occur south of the Sahara, may have also been caused by this shift to greater aridity, especially considering that extensive desert conditions in the Sahara did not occur until the Late Pliocene, around 2.8 MYA 747677.

If mustelid diversification was promoted by climatically driven paleoenvironmental changes, then synchronous patterns and tempos of diversification should also be recorded in the evolutionary histories of other, unrelated groups. A number of recent molecular phylogenetic studies on various groups of mammals and other vertebrates employing multiple loci and relaxed molecular clock methods provide corroboration for this hypothesis. For example, squirrels (Sciuridae 32), cats (Felidae 78), rabbits and hares (Leporidae 79), deer (Cervidae 80) and woodpeckers (Picinae 81) each show one or more episodes of rapid diversification that are roughly contemporaneous with one or both of those observed in mustelids (Figure 5). Taking cats as an example, seven of the eight primary lineages of felids radiated in the early part of the Late Miocene (10.8-6.2 MYA) whereas genera and species that comprise the eight lineages largely radiated during the Pliocene 78. We also note that divergence events from a number of independent lineages within the Mustelidae are synchronous as well. For example, using the root age and crown age prior combination of 28.5 MYA and 24 MYA as reference, nodes 11, 14, 19, 22, 31 and 38 all codiversify around 2.8 MYA (Figure 2 and Table 2). Similarly, nodes 9, 28 and 37 overlap in their divergence time estimates of around 4.8 MYA (Table 2). While ecological circumstances obviously differ from group to group, such congruence in patterns and tempos of diversification supports the idea that a common cause, namely, large-scale changes in past environments, has shaped the phylogenetic histories among disparate groups of organisms as well as independent lineages of mustelids.

Our results have important implications for theories about the mode and tempo of adaptive radiations. First, the scenario of diversification across different trophic groups we have outlined above is consistent with the model of cascading radiation proposed by Stanley 82, in which diversification of one group in a trophic cascade (e.g. producers such as grasses and taiga) promotes diversification in a second group (e.g. primary consumers such as rodents), which in turn promotes diversification in a third group (secondary consumers such as mustelids). This model could be tested more rigorously using dated phylogenies of multiple groups representing different trophic levels. Second, our finding that the two primary bursts of cladogenesis in the evolutionary history of extant mustelids (and perhaps in the other groups cited above) coincided with periods of climatically mediated environmental changes provides support for models that show that evolution may be greatly accelerated in temporally and spatially changing environments 83. During these times, new niches are created or former niches are reorganized, providing new ecological opportunities that may foster rapid speciation and thus, diversification 284 (see also 85). The fact that multiple groups show contemporaneous periods of rapid cladogenesis (e.g. mustelids, felids and leporids during the Pliocene) suggests that mainland environments undergoing environmental changes may function like newly colonized island archipelagoes in promoting diversification.

The two-state and four-state biogeographical reconstructions indicate that the vast majority of the modern diversification of mustelids has occurred in the Old World (Figure 3), specifically in Eurasia (Figure 4). In fact, Eurasia was unambiguously reconstructed as the ancestral area for nearly every ancestral node in the four-state analysis (Figure 4). These results are consistent with two other observations that suggest Eurasia has been the center of mustelid diversification: (1) Eurasia contains the majority of extant species, with 34 of the 59 known species either exclusively endemic to or having part of their distribution on this continent; and (2) the earliest fossils of extant lineages or those associated with the ancestors of extant lineages are often found in Eurasia 23. Nodes located near the base of the tree tend to be those where likelihood ratios are not significant (and, thus, ancestral reconstruction is inferred as equivocal), such as the root node and the node joining Taxidea as sister to the remaining taxa of mustelids in the two-state analysis (see Figure 3 and Additional file 3). Such results are not uncommon in likelihood or other model-based approaches to ancestral state reconstruction because the degree of uncertainty associated with reconstruction increases with time 86. Nonetheless, proportional likelihood values for these nodes still favor the Old World as the ancestral area in the two-state analysis (see Figure 3 and Additional file 3). Ancestral reconstructions were robust when Enhydra lutris, Gulo gulo and Mellivora capensis were coded for alternative states (results not shown; see methods).

In contrast to the extensive in situ diversification that has taken place in Eurasia, mustelid faunas of Africa and the New World are largely comprised of genera or species that repeatedly colonized these regions from Eurasia. The mustelid fauna of Africa contains eight species, seven of which are included in our taxon sampling (the Egyptian weasel, Mustela subpalmata, was not sampled). Of these seven species, five are derived from separate colonizations from Eurasia (Aonyx capensis, Hydrictis maculicollis, Mustela putorius, the Ictonyx + Poecilogale lineage and Mellivora capensis) whereas two are derived from in situ speciation events (Ictonyx striatus and P. albinucha); see Figure 4. Similarly, our reconstructions show that nine separate dispersal events from Eurasia and only one in situ speciation event accounts for the diversity of mustelids that are either endemic to North America (e.g. Lontra canadensis, M. nigripes, Martes pennanti) or have part of their distribution there (e.g. Holarctic species such as Gulo gulo, Mustela erminea and M. nivalis). Genera and species of mustelids found in South America today are largely descended from North American immigrants that arrived as part of the Great American Interchange following the rise of the Panamanian isthmus, 3.0-2.5 MYA 238788. Such a relationship is clearly indicated for the clade of New World otters in which L. canadensis is sister to L. felina + L. longicaudis, with the latter two species found in Central and/or South America (Figure 4). Moreover, this split (node 14) is estimated to have occurred 2.8–3.4 MYA (95% HPD: 1.6–5.2 MYA), which overlaps well with timing of the formation of the Panamanian land bridge. The long-tailed weasel, M. frenata, ranges from North America to northern South America 89 and two species of weasels (M. africana and M. felipei, not sampled here) are endemic to South America. Fossil evidence clearly indicates that Mustela colonized South America from the north, apparently well after the Panamanian isthmus was in place 2390. Our results show that Pteronura, Galictis and Eira dispersed separately into South America, with Eurasia reconstructed as the continent of origin for each genus (Figure 4). These results are anomalous because these genera have been allied with extinct taxa from North America, suggesting a more proximate origin for these lineages 9192. For example, Pteronura may be related to the extinct genus Satherium from the Pliocene of North America 91. However, paleontological studies suggest that the ultimate ancestry of these extinct taxa lies in Eurasia 9293.

A combination of evidence from the fossil record and divergence times from our phylogeny indicates that the mustelid faunas of Africa, North America and South America have been assembled gradually over time. For example, fossil evidence suggests mustelids colonized the New World across Beringia during different intervals when the land bridge between Eurasia and North America was open. Multiple genera of mustelids entered North America during the Late Miocene (~11.2-5.3 MYA 1923), prior to the first opening of the Bering Strait 5.4–5.5 MYA, which severed the route across Beringia 9495. Many genera that colonized North America during the Late Miocene or earliest Pliocene became extinct 1920. Nonetheless, among the genera that may have been included in this wave of dispersal were the earliest representatives of Lutra (which may represent Lontra, given that New World river otters have been reclassified into Lontra) and Mustela, both of which are first recorded in North America from the Late Miocene to Early Pliocene (~5.9-4.6 MYA 19). These taxa may have been the forerunners of modern species of Lontra and Mustela/Neovison found in North and South America today. The Late Miocene divergence time of the splits leading to Lontra and Neovison + M. frenata (nodes 8 and 16 in Figure 2) are consistent with this possibility. Two extinct genera of American badgers, Chamitataxus and Pliotaxidea, are recorded from the Late Miocene, around 7.3 MYA and 6.5 MYA, respectively 9697. Pliotaxidea has been shown to be sister group to Taxidea based on morphological evidence 98, thereby suggesting that the lineage leading to Taxidea arrived in North America before the opening of the Bering Strait. Meline badgers (Arctonyx and Meles) are presently found only in the Old World. However, recent discovery of Late Miocene to Early Pliocene fossils of meline badgers at two different sites in North America 99100 indicates that this lineage had also immigrated into North America and was a component of the New World mustelid fauna.

Following these earlier dispersal events, fossil evidence indicates that Mustela erminea, M. nigripes, M. nivalis and Martes americana later entered North America during the Pleistocene 26102. The molecular divergence time for M. nigripes (node 25), around 0.6 MYA (95% HPD: 0.3–1.1 MYA; Table 2), for example, supports a Pleistocene dispersal scenario for this species. Although fossil records of mustelids in Africa and South America are less well known than those of the northern continents, first appearance datums of both extant and extinct genera of mustelids nevertheless suggest that these regions also were colonized through successive dispersal events 23102103104105. Among the extinct genera known from Late Miocene deposits in East Africa that may have arisen from Eurasian immigrants are the gigantic and cat-like Ekorus, the largest mustelid discovered thus far, and an otter, Vishnuonyx 105. In contrast, the earliest known remains of Ictonyx striatus are from the Pleistocene 106, suggesting a later immigration into Africa.

Most studies of modern-day mustelid communities (where multiple species coexist in a single area of a geographic region) have focused on recent ecological factors, such as partitioning of resources (food or space) via competition, to explain coexistence of species within these communities 29107108109. As these studies have operated on an ecological timeframe, they implicitly assume that ecological differences among species have evolved relatively recently. However, several observations derived from our biogeographical and phylogenetic results suggest that history (via phylogeny) also has been an important component in the structure of mustelid assemblages on different continents.

First, these assemblages are largely composed of species belonging to different clades or lineages that differ significantly in their ecomorphology with regards to diet and locomotor mode 110111. For example, seven species of mustelids are sympatric on the British Isles in Eurasia, with three species of weasels, (Mustela) and one species each of mink (Neovison, introduced), marten (Martes), badger (Meles) and otter (Lutra) 29. Weasels, martens, badgers and otters obviously comprise a phylogenetically and ecologically heterogeneous set of species, yet these fundamental differences may indeed facilitate coexistence of these species within this community. Even the three species of weasels found in this community (M. erminea, M. nivalis and M. putorius) are not closely related (Figure 1), suggesting that resource partitioning among these species may also be in part determined by historical causes.

Second, studies of other vertebrate communities (e.g. lizards) have demonstrated that competition and divergence in ecological traits varies according to the degree of phylogenetic relatedness among species, such that, for example, competition is expected to be low among distantly related species but high among closely related species 112113. Divergence in ecological traits should therefore be greatest between closely related species, especially sister species occupying the same area. Within mustelids, there are several instances where sister taxa that are broadly sympatric within a region often exhibit pronounced ecological differences. In Africa, Poecilogale albinucha preys almost exclusively on rodents whereas Ictonyx striatus has a more generalized diet that includes invertebrate prey 3637. Similarly, Mustela frenata and Neovison vison overlap extensively in North America, but differ significantly in several aspects of their ecologies 89114. The sister taxa Aonyx cinerea and Lutrogale perspicillata co-occur in parts of south and southeast Eurasia with the former species feeding largely on freshwater crabs and the latter feeding mostly on fish 115116.

Third, the fossil record along with molecular dating results indicate that continental assemblages of mustelids have been built up gradually over time (see the above discussion). This suggests that species interactions with these communities are temporally mosaic, with some interactions being ancient while others are more recent. We certainly do not suggest that present-day ecological factors (e.g. interspecific competition for food and/or space, predation) operating over ecological time scales have not been important in determining the structure of modern-day mustelid communities. Instead, we simply suggest that historical factors should also be considered and that such a perspective has been lacking in most previous studies of mustelid community ecology. As other researchers have pointed out, both historical factors and more recent ecological processes contribute to the structure of modern day communities 113117.

We obtained tissue samples from mustelid taxa representing 21 of 22 putative genera and 43 of 58 extant putative species, following the classification of Wozencraft 7; (see Additional file 4). Samples were unavailable from Lyncodon patagonicus, seven species of Mustela, two species of Meles, two species of Melogale and one species each of Lontra, Lutra and Martes. Multiple lines of evidence indicate that Procyonidae is the sister group of Mustelidae 1016. Therefore, two species of Procyonidae, Procyon lotor and Bassariscus astutus, were used as outgroups to root the mustelid tree.

Total genomic DNA was extracted from hair or tissue samples using phenol chloroform, followed by ethanol precipitation 118 or using a QIAamp DNA Mini Kit (Qiagen, Valencia, CA). Twenty-one nuclear gene segments were amplified using published primers (see Additional file 5) and the complete mitochondrial cytochrome b (CYTb) gene was amplified with primers as described in Koepfli and Wayne 13. Polymerase chain reaction (PCR) was carried out in MWG-Biotech Primus 96 Plus thermal cyclers with the following conditions: 28–30 cycles of 94°C for 30 s; 50–56°C for 30 s; 72°C for 45 s; and one cycle of 72°C for 5 min. Each 50 μl reaction contained 35.7 μl sterile double-distilled water, 5 μl 10× PCR buffer, 5 μl of 25 mM MgCl₂, 1 μl of 10 mM dNTP mix, 1 μl of both 25 pM/μl forward and reverse primers, 0.3 μl Taq polymerase (Sigma-Aldrich, St Louis, MO) and 1 μl of 0.1–1 μg genomic DNA. All PCRs included a negative control (no DNA). PCR products were electrophoresed in and excised from 1% agarose/Tris-acetic acid-EDTA gels and purified using an Ultra Clean Kit (MoBio Laboratories, Solana Beach, CA). PCR products were then cycle sequenced using the original amplification primers and either the CEQ Dye Terminator Cycle Sequencing Quick Start Kit (Beckman Coulter, Fullerton, CA) or BigDye Terminator v3.1 Cycle Sequencing Kit (Applied Biosystems, Foster City, CA). Sequencing reactions were run through either a CEQ 2000 XL DNA Analysis System or an Applied Biosystems 3730 DNA Analyzer. Sequence chromatograms were checked for accuracy and edited using Sequencher 3.1 (Gene Codes Corporation, Ann Arbor, MI).

For several species, DNA extracts from hair samples did not yield a sufficient amount of DNA for direct amplification of nuclear gene segments. We therefore first whole-genome amplified these samples using the GenomiPhi V2 DNA Amplification Kit (Amersham Biosciences, Little Chalfont, UK) and then proceeded with regular PCR protocols (as above). To ensure that the whole-genome amplification process had not introduced any errors into our target sequences, we amplified the whole genome from the same sample a second time and then amplified and sequenced several nuclear loci to compare sequences from the two samples (all were identical).

Mustela altaica was represented by only three of the 22 gene segments: CYTb, RAG1 and APOB (exon 26). Sequences for these segments were downloaded from Genbank and were from studies by Sato et al 1516 and Kurose et al 33, thereby increasing the ingroup to 44 taxa. RAG1 sequences for Gulo gulo, Martes flavigula, M. foina, M. martes, M. zibellina, Melogale moschata, Mustela altaica and M. erminea were downloaded from Genbank from the study by Sato et al 15. All new sequences were deposited in Genbank and accession numbers for all sequences are presented in Additional file 6.

We were unable to obtain sequences for one or more gene segments from nine of the 46 species sampled. Species and the segments they are missing are: Galictis cuja (BRCA1 [fragment 2]); Ictonyx libyca (ATP7A); Martes zibellina (PLCB4); Mellivora capensis (ADORA3); Melogale personata (APOB [exon 29], BRCA1 [fragment 1], BRCA1 [fragment 2], CHRNA1, FES, GHR, GLB1, GNAT1, INHBA, RHO1, TMEM20, WT1); Mustela nudipes (BRCA1 [fragment 2], COL10A1, FES, INHBA, TMEM20); M. strigidorsa (BRCA1 [fragment 2], COL10A1, TMEM20); Procyon lotor (TMEM20); Vormela peregusna (COL10A1, FES, GLB1). Question marks were used to represent missing sequences. Despite the amount of missing data for species such as Melogale personata and Mustela altaica, studies have shown that phylogenetic information content of included data for a taxon is more important in achieving phylogenetic accuracy than the amount of missing data per se, especially in the context of a supermatrix analysis 119120.

Gene segments were concatenated into a supermatrix of 11,929 bp, including insertions and deletions (indels). Phylogenetic trees were estimated using MP, ML and BI. Indels were coded as missing for BI and ML analyses (11,929 bp) but were recoded as present or absent (1 or 0), regardless of length, to utilize their potential phylogenetic signal for MP analyses (11,789 bp) 121. PAUP* 4.0b10 122 was used to reconstruct MP trees. Characters were equally weighted in heuristic searches using 1,000 random stepwise additions, with one tree held at each step during stepwise addition, tree-bisection-reconnection branch swapping, steepest descent option not in effect, no upper bound for MaxTrees and MulTrees option in effect. Clade support was evaluated by bootstrapping, using 3,000 pseudoreplicates and the same heuristic search conditions as described above except only 100 random stepwise additions were used.

The GTR+I+G model was selected as the best-fitting model of DNA substitution for the 11,929 bp data set, using the Akaike information criterion (AIC) as implemented in Modeltest v3.7 123. Under this model, ML heuristic searches were conducted using a hill-climbing algorithm and a genetic algorithm as implemented in the programs TREEFINDER 124 and GARLI 125126, respectively. For GARLI, we used a random starting tree and default settings for the components of the genetic algorithm. Identical topologies and similar log-likelihood scores were obtained for three separate runs with each program. For both methods of ML analysis, 1,000 bootstrap pseudoreplicates were used to assess the support for the ML topology.

We used MrBayes v3.1.2 127 for Metropolis-coupled MCMC BI of phylogeny. We performed MCMC runs under the GTR+I+G model of DNA substitution, selected by MrModelTest v2.2 128 using the AIC. MCMC analyses were performed in which model parameters were linked (uniform model) or unlinked (partitioned model) among the 22 gene segments of the concatenated data set. Two simultaneous independent runs of one cold and three heated MCMC chains and each starting with a different random tree were conducted for 5 × 10⁶ generations, sampling trees every 500 generations. To ensure consistency of results, we ran analyses for both models a second time (four independent runs for both uniform and partitioned models). The following set of priors were used in all analyses: Dirichlet priors for six substitution rates of the GTR model (1, 1, 1, 1, 1, 1); a Dirichlet prior for base frequencies (1, 1, 1, 1); a uniform prior for the proportion of invariant sites (0, 1); a uniform distribution prior for the shape parameter of the gamma distribution of rate heterogeneity among sites (0, 200); all topologies equally probable; and unconstrained branch lengths with an exponential probability density. Potential scale reduction factors (PSRFs) of 1.00 and an average standard deviation of split frequencies for both simultaneous runs of less than 0.01 indicated that runs had converged on a stationary distribution. In addition, using Tracer 1.3 129, tracer plots and effective sample size values over 200 for estimates of the posterior distribution of tree likelihood and model parameters also indicated that convergence had been reached and that MCMC chains had mixed well. For each independent run, the first 3,000 trees were discarded as burn-in, leaving 14,002 trees used to construct a 50% majority-rule consensus tree. Internodes with posterior probability values of ≥0.95 were considered well supported.

We reconstructed ancestral areas of mustelids using the likelihood reconstruction method 86130 implemented in Mesquite v1.12 131. Taxa were coded into one of two categorical characters, 0 = Old World (Africa, Eurasia) or 1 = New World (North America, South America) or one of four categorical characters, 0 = Eurasia, 1 = North America, 2 = Africa or 3 = South America. Three taxa have Holarctic distributions, Gulo gulo, Mustela erminea and M. nivalis, while a fourth taxon, Enhydra lutris, is distributed along coastal waters of the eastern and northern Pacific Ocean. As these four taxa have distributions that span both Old and New Worlds (two state) or Eurasia and North America (four state), they would be coded as polymorphic (0, 1). However, polymorphic characters cannot be used with the likelihood reconstruction method of Mesquite, so we used fossil evidence to assign the four taxa to one of the categorical characters. Mustela erminea and M. nivalis were coded as 0 = Old World or Eurasia since the earliest fossil remains of these taxa are found in Eurasia 132133. The earliest fossil evidence for the wolverine, G. gulo, is found in North America, although Eurasian fossils are almost contemporaneous 134135. Gulo, however, is either related to or descended from Plesiogulo, which originated in Asia in the Late Miocene 136. Therefore, we coded Gulo as 0 = Old World or Eurasia. The earliest fossil remains for Enhydra are found in Pleistocene deposits of North America 93137. Cladistic analyses suggest that Enhydra shares ancestry with the extinct Enhydritherium, which immigrated to North America from Eurasia in the Late Miocene 138. Although there is debate about the exact area of origin of Enhydra within North America 93139, the earliest fossils leading to this lineage are of Old World origin 93 and accordingly, we coded this taxon as 0 = Old World or Eurasia. In addition, M. frenata is distributed from North America to South America, with the earliest fossil remains found in the former region 89. Consequently, we coded this species 1 = North America in the four state analyses. Lastly, fossil evidence suggests that Mellivora capensis originated in Africa, although the species also ranges into Eurasia 103. Therefore, this species was coded 3 = Africa in the four state analysis. Taxa whose exact geographic origins were uncertain (i.e. E. lutris, G. gulo and M. capensis) were coded by the alternative state in separate analyses in order to investigate robustness of the biogeographic reconstructions (e.g. G. gulo was coded 1 = New World or North America in two-state and four-state analyses, respectively).

For reconstruction of ancestral areas using likelihood, we used the Markov k-state one-parameter model (Mk1) 140, which assumes a single rate of transition between two character states. The rooted topology and branch lengths generated from the ML analyses were used to trace characters. We used the likelihood-ratio test to determine the best estimate of the reconstructed character state at each node, setting the likelihood decision threshold at 2.0. If log-likelihoods of two states differed by 2.0 or more, the state with the lower negative log-likelihood was accepted as the best estimate 130. Character state reconstructions were considered ambiguous at nodes where log-likelihoods differed by less than 2.0.

We estimated divergence times of splits using the Bayesian relaxed phylogenetic approach implemented in BEAST v1.4.2 52141. We assumed a GTR+I+G model of DNA substitution with four rate categories. Uniform priors were employed for GTR substitution parameters (0, 100), gamma shape parameter (0, 100) and proportion of invariant sites parameter (0, 1). The uncorrelated lognormal relaxed molecular clock model was used to estimate substitution rates for all nodes in the tree, with uniform priors on the mean (0, 100) and standard deviation (0, 10) of this clock model. We employed the Yule process of speciation as the tree prior and a UPGMA tree to construct a starting tree, with the ingroup assumed to be monophyletic with respect to the outgroup. Eight fossil calibration points were applied as priors to constrain the age of the following nodes (as numbered in Figures 1 and 2): (i) node 1, 24.0 MYA as the minimum age for crown Mustelidae, based on the earliest known mustelid, Plesictis, from the Late Oligocene of Europe 1418 (but see 25); (ii) node 2, 10.0 MYA as the minimum age for origin of Mellivora, from Late Miocene deposits in South Africa 103; (iii) node 10, 3.6 MYA as the minimum age for origin of Lutra, based on fossils of Lutra affinis from the Early Pliocene of Europe 93; (iv) node 11, 1.0 MYA as the minimum age for the origin of Aonyx, based on Pleistocene fossils of this species from Africa 104; (v) node 16, 5.3 MYA as the minimum age for origin of Mustela (including Neovison), based on fossils from a number of different species of Mustela that appear during the Late Miocene throughout Eurasia 142; (vi) node 18, 1.8 MYA as the minimum age for the earliest fossil remains of M. erminea 132; (vii) node 28, 1.8 MYA as the minimum age for origin of Vormela, based on Pleistocene fossil remains of V. petenyii from Europe 143; and (viii) node 38, 3.3 MYA as the minimum age for origin of the subgenus Martes, based on fossils of Martes wenzensis from the Late Pliocene of Europe 73 (see also 14). We conducted additional dating analyses by using minimum fossil constraints (ii)–(viii), but changed the age of crown Mustelidae to a maximum of 28.5 MYA, using the first appearance of the stem taxon Pseudobassaris (considered the earliest known taxon of the Procyonidae) as the earliest age of the Mustelidae-Procyonidae split 144 (see also 14). All fossil constraint priors were set as means of a normal distribution, with a standard deviation of 1.0 MYA. We set the mean of the normal distribution of the root height prior to 24 MYA (assuming Plesictis as the earliest fossil mustelid), with a standard deviation of 1.0 MYA. To assess the influence of the root height prior on resulting node ages, we also conducted runs in which age of the root height prior was increased to 28.5 MYA and 33.7 MYA, representing minimum and maximum ages for the Early Oligocene. These ages correspond to the approximate times for the Mustelidae-Procyonidae split 22144. Three independent MCMC runs of 10,000,000 steps were performed for each combination of crown group age prior (24 MYA, 28.5 MYA) and root height prior (24 MYA, 28.5 MYA, 33.7 MYA), with parameters logged every 1,000 steps. The Auto Optimize Operators function was enabled to maximize efficiency of MCMC runs. Three independent MCMC runs for each analysis were combined to estimate the posterior distribution of the substitution model and tree model parameters, as well as node ages. Analyses of these parameters in Tracer 1.3 129 suggested that the number of MCMC steps was more than adequate, with effective sample sizes of all parameters often exceeding 1,000 and Tracer plots showing strong equilibrium after discarding burn-in.