Cichlid samples were obtained from local animal dealers in Japan. We combined these new mitogenomic data with 48 previously published sequences from the DDBJ/EMBL/GenBank nucleotide sequence database. The 10 cichlid taxa that we analyzed (Table 1) cover species from major Gondwana-origin landmasses. In addition, we chose 31 other teleosts, nine basal actinopterygians, and two sarcopterygians. Two sharks were sampled as an outgroup to root the tree. Additional file 1 contains a complete list of the sampled taxa, along with the database accession numbers of their mitogenomic sequences.

Fish samples were excised from live or dead specimens of each species and immediately preserved in 99.5% ethanol. Total genomic DNA was extracted from muscle, liver, and/or fin clips using a DNeasy tissue kit (Qiagen) or a DNAzol Reagent (Invitrogen), following manufacturer protocols. The mtDNA of each species was amplified using a long-PCR technique with LA-Taq (Takara). Seven fish-versatile primers for long PCR (S-LA-16S-L, L2508-16S, L12321-Leu, H12293-Leu, H15149-CYB, H1065-12S, and S-LA-16S-H 212223242526) and the two cichlid-specific primers cichlid-LA-16SH (5'-TTGCGCTACCTTTGCACGGTCAAAATACCG-3') and cichlid-LA-16SL (5'-CGGAGTAATCCAGGTCAGTTTCTATCTATG-3') were used in various combinations to amplify regions covering the entire mtDNA in one or two reactions. The long-PCR products were used as templates for subsequent short PCR.

Over 100 fish-versatile PCR primers 21222324252627 and 18 taxon-specific primers (Additional file 2) were used in various combinations to amplify contiguous, overlapping segments of the entire mtDNA for each of the six new cichlid species. The long PCR and subsequent short PCRs were performed as described previously 2128. The short-PCR reactions were performed using the GeneAmp PCR System 9700 (Applied Biosystems) and Ex Taq DNA polymerase (Takara).

Double-stranded PCR products, treated with ExoSAP-IT (USB) to inactivate remaining primers and dNTPs, were directly used for the cycle sequencing reaction, using dye-labeled terminators (Applied Biosystems) with amplification primers and appropriate internal primers. Labeled fragments were analyzed on Model 3100 and Model 377 DNA sequencers (Applied Biosystems).

The DNA sequences obtained were edited and analyzed using EditView 1.0.1, AutoAssembler 2.1 (Applied Biosystems) and DNASIS 3.2 (Hitachi Software Engineering Co. Ltd.). Individual gene sequences were identified and aligned with their counterparts in 48 previously published mitogenomes. Amino acid sequences were used to align protein-coding genes, and standard secondary structure models for vertebrate mitochondrial tRNAs 29 were consulted for the alignment of tRNA genes. The 12S and 16S rRNA sequences were initially aligned using clustalX v. 1.83 30 with default gap penalties and subsequently adjusted by eye using MacClade 4.08 31.

The ND6 gene was excluded from the phylogenetic analyses because of its heterogeneous base composition and consistently poor phylogenetic performance 22. The control region was also excluded because positional homology was not confidently established among such distantly-related species. The third codon positions of protein genes were excluded because of their extremely accelerated rates of change that may cause high levels of homoplasy. After the exclusion of unalignable parts in the loop regions of tRNA genes, as well as the 5' and/or 3' end regions of protein genes, all gene sequences were concatenated to produce 10,034-bp sites (6962, 1402, and 1670 positions for protein-coding, tRNA, and rRNA genes, respectively) for phylogenetic analyses.

Phylogenetic trees were reconstructed using partitioned Bayesian and maximum likelihood analyses. Partitioned Bayesian phylogenetic analyses were performed using MrBayes 3.1.2 32. We set four partitions (first codon, second codon, tRNA, and rRNA positions). The general time-reversible model, with some sites assumed to be invariable and variable sites assumed to follow a discrete gamma distribution (GTR + I + Γ; 33), was selected as the best-fit model of nucleotide substitution by MrModeltest 2.2 http://www.abc.se/~nylander/34. The Markov chain Monte Carlo (MCMC) process was set so that four chains (three heated and one cold) ran simultaneously. We ran the program for 3,000,000 metropolis-coupled MCMC generations on each analysis, with tree sampling every 100 generations and burn-in after 10,000 trees.

Partitioned maximum likelihood (ML) analyses were performed with RAxML ver. 7.0.3 35, a program implementing a novel, rapid-hill-climbing algorithm. For each dataset, a rapid bootstrap analysis and search for the best-scoring ML tree were conducted in one single program run, with the GTR + I + Γ nucleotide substitution model. The rapid bootstrap analyses were conducted with 1000 replications, with four threads running in parallel.

Statistical evaluation of alternative phylogenetic hypotheses was done using TREE- PUZZLE 5.2 36, using the two-sided Kishino and Hasegawa (KH) 37 test, the Shimodaira and Hasegawa (SH) 38 test, and Bayes factors 3940. We used the GTR + I + Γ model and its parameters optimized by MrModeltest 2.2.

For the divergence time estimation, multidistribute program 41 was used by assuming a topological relationship thus obtained, but without assuming the molecular clock (i.e., by allowing heterogeneity in molecular evolutionary rate along branches). Upper and/or lower time constraints at selected nodes were set for the Bayesian MCMC processes to estimate divergence times (including means and 95% credibility ranges) and relative rates at ingroup nodes. We set the partitioning as described above and first used PAML 42 to optimize the parameters of model F84 and the gamma distribution for eight categories to account for site heterogeneity. Estbranches and multidivtime programs were then used to estimate divergence times. We used 21 fossil-based time constraints assignable to diverse teleostean lineages (Table 2).

We determined complete or nearly complete mtDNA nucleotide sequences for six new cichlids from Africa, South America, Madagascar, and Indo/Sri Lanka (Table 1). The sizes of these genomes ranged from 16,457 to 16,556 bp, including approximately 800 bp in the control region. Tylochromis polylepis alone appears to have a somewhat longer control region (approximately 1200 bp) although the exact sequence of the region was unable to be determined because of the long poly-T sequences within the region. We also analyzed the previously published mitogenomic sequences of four cichlid species (Table 1). Oreochromis mossambicus (accession no. AY597335) was not included because a congeneric taxon (Oreochromis sp.) sequenced by Mabuchi et al. 43 had already been sampled.

All 37 genes encoding two rRNAs, 22 tRNAs, and 13 proteins were identified in these 10 cichlid mitogenomes, basically in the same order and orientation found for most other vertebrates. Transfer RNA genes could be folded into secondary structures typical of vertebrate mitochondrial tRNA 29. The base composition of cichlid mitogenomes was skewed (data not shown) similarly to those of other vertebrates 44.

Figure 1 shows the phylogenetic relationships inferred from the Bayesian analysis among the 52 bony fishes, estimated with two sharks as an outgroup. The tree topology was identical to that obtained by the partitioned ML analysis (data not shown). These bony fish taxa included two sarcopterygians (coelacanth and lungfish), nine basal actinopterygians (polypterids, acipenseriforms, lepisosteids, and amiid), and 41 teleosts, including 10 cichlids. The phylogenetic relationships obtained for non-cichlid taxa (Fig. 1) were largely consistent with those from previous mitogenomic studies 284345, except for a difference in the sister group of holosteans (lepisosteids and amiid).

Although Inoue et al. 28 suggested that the "Ancient Fish Clade" unites acipenserids, lepisosteids, and amiid, our phylogenetic analysis supports the neopterygian clade (lepisosteids + amiid + teleosts), in agreement with an analysis of nuclear DNA sequences 46 and morphological characters 47. Relationships between the basal actinopterygians and teleosts were not stable against changes in taxonomic representations and the genes used and varied between the two hypotheses (data not shown). We tentatively assumed the neopterygian relationship for subsequent analyses because this was consistent in both morphological and molecular (based on mitochondrial and nuclear sequences) analyses. However, we also conducted analyses to evaluate how our major conclusions in dating depend on the two alternative phylogenetic relationships (Table 3).

In terms of the relationships among 20 percomorphs containing 14 labroids (two labrids, two pomacentrids, and 10 cichlids), we reconfirmed the polyphyly of Labroidei 43 whereby labrids (designated Labroidei 1 in Fig. 1) and cichlids + pomacentirids (Labroidei 2) appear in separate lineages of teleosts. The non-monophyly of the labroid taxa was supported by a number of nodes with 100% posterior probability and 100% bootstrap values (Fig. 1).

Among the 10 cichlid taxa that we used, four were from Africa, two from South America, three from Madagascar, and one from Indo/Sri Lanka. The tree (Fig. 1) supports the monophyly of Cichlidae and two other continental groups from Africa and South America. Four basal taxa from Madagascar and Indo/Sri Lanka are not monophyletic, and two (Paretroplus from Madagascar and Etroplus from Indo/Sri Lanka) corresponding to Etroplinae sensu Sparks and Smith 16 form a sister group to all other cichlids. The other two Malagasy taxa (Paratilapia and Ptychochromoides), corresponding to Ptychochrominae sensu Sparks and Smith 16, form a sister group to the African + Neotropical clade. These results are consistent with previous molecular studies that used a few mitochondrial or nuclear gene sequences 14151648, as well as morphological studies 13.

However, these previous studies did not fully evaluate the statistical significance in rejecting alternative hypotheses of cichlid relationships. We conducted KH and SH tests, as well as a test using Bayes factor. Based on these tests, alternative hypotheses assuming the monophyly of Malagasy + Indo/Sri Lankan cichlids (constraint 1), Old World cichlids (constraint 2), and African + Indo/Sri Lankan cichlids (constraint 3) are all very unlikely (Table 4). These results provide statistical support for the paraphyletic assemblage of the Malagasy + Indo/Sri Lankan taxa to the African + Neotropical clade.

If Cichlidae originated in Cenozoic Africa and migrated into South America, Madagascar, and India via saltwater dispersal 1949, Malagasy/Indo Sri Lankan and/or Neotropical taxa would probably be nested in the African clade, and alternative relationships (e.g., those corresponding to constraints 2 and 3) would likely appear. However, these relationships were not found, thus supporting the vicariant divergence scenario 131418, at least from a topological standpoint.

We conducted divergence time estimation among 54 bony fishes, including 10 cichlids (Fig. 2). Twenty-one time constraints based on extensive fossil evidence for bony fishes (Table 2) were used. Following the advice of Benton and Donoghue 2 to set fossil-based time constraints as hard lower boundaries and soft upper boundaries, we chose older values for upper boundaries. We estimated the divergence between African + Neotropical cichlids and Malagasy + Indo/Sri Lankan (ptychochrominae) cichlids to be approximately 96 MYA (78–115 MYA at 95% credibility). The divergences of African vs. Neotropical cichlids and Malagasy vs. Indo/Sri Lankan cichlids within the Etroplinae were estimated to be approximately 89 MYA (72–108 MYA) and 87 MYA (69–106 MYA), respectively.

We then compared the estimated divergence times among cichlids and the probable times of continental fragmentation based on geological evidence. The divergence time between Malagasy and Indo/Sri Lankan taxa within Etroplinae (~87 MYA: 69–106 MYA) is very close to the time of separation between Madagascar and India (85–95 MYA) 5051. The divergence time estimated between African and Neotropical clades (~89 MYA: 72–108 MYA) is also close to the time of separation between African and South American landmasses (~100 MYA) 5051. The divergence time between African + Neotropical cichlids and Malagasy ptychochrominae cichlids (~96 MYA: 78–115 MYA) appears to be somewhat more recent than the time generally accepted for the complete separation of the Indo-Madagascar landmass from Gondwanaland (120–130 MYA) 5051. However, some studies 52 have postulated an extended connection between India and Antarctica by approximately 112 MYA, which is within the 95% credibility range for the African/Neotropical vs. ptychochrominae cichlid divergence. Taken together, these results are consistent with the vicariant divergence of continental cichlid groups during Cretaceous times and argue against their Cenozoic dispersal.

Vences et al. 19 calibrated a molecular clock for cichlids that assumed that the divergence time of the most basal endemic lineages in East African Rift lakes (e.g., Tanganyika) corresponds to the geological estimate of the age of the lakes. These estimated divergence times between continental cichlid clades were all in the Cenozoic (rather than the Mesozoic, as we demonstrate in Fig. 2) and supported the hypothesis of long-distance Cenozoic transmarine dispersal of cichlids. This view of the Cenozoic (or latest Cretaceous) origin and transmarine dispersal of cichlids has also been supported by some biogeographers 49 because it is consistent with cichlid fossil records, which first occur in South America and Africa in the Eocene 2053. However, the clock-based dating procedures of Vences et al. 19 present some problems. The strict molecular clock may not hold for all cichlid lineages 15, and the premise that the oldest endemic cichlid divergence is synchronized with the formation of the lakes may not be valid. Some lineages that had diverged outside the lake may have immigrated in parallel 7. In addition, there is no definitive, geologically based time estimate for the formation of the lakes.

More recently, Genner et al. 7 used two mitochondrial (cytochrome b and 16S rRNA) and one nuclear (TMO-4C4) gene fragments to estimate the divergence times among cichlids. When the cichlid divergence by Gondwanan vicariance was assumed, the resultant divergence times were more consistent with those estimated with time constraints from previous paleontological and molecular studies 254555657 than when the Cenozoic cichlid divergence was assumed based on fossil records.

Although we concur on the Gondwanan origin and vicariant divergence of cichlids, Genner et al. 7 evaluated this biogeographic hypothesis somewhat indirectly, in that the fitness of estimated times of cichlid divergences to those obtained with time constraints from previous studies was qualitatively compared between alternative assumptions on cichlid biogeography. We evaluated cichlid divergence times more directly by using longer mitogenomic sequence data and dozens of non-cichlid taxa, allowing us to set many time constraints purely from the paleontological data and providing additional evidence for an ancient cichlid divergence on Gondwanaland, despite the general paucity of the Mesozoic and Cenozoic paleontological record on bony fishes.

In Figure 3, minimum time constraints based on fossil records (see Table 2) are plotted against molecular time estimates of the corresponding divergences (values taken from Fig. 2). In this figure, minimum age estimates of Gondwanan fragmentations are also plotted against the corresponding molecular time estimates of continental cichlid groups. It should be noted here that the latter data points reflecting Gondwanan fragmentation history (closed triangles) are plotted well on the line of 1:1 relationship whereas most of the data points reflecting fossil records (closed circles) are considerably below the line of the 1:1 relationship. This pattern suggests that Gondwana fragmentation history that is congruent with the cichlid phylogeny can be effective time constraints better than most of the Mesozoic and Cenozoic fossil records used here.

Among the fossil data points, four data points in the Paleozoic show a fairly good 1:1 relationship, whereas other points mostly in the Mesozoic are considerably below the line of 1:1 relationship. This might mean that the Mesozoic fossils do not really represent the oldest fossil for the corresponding lineages whereas this is not the case for older Paleozoic lineages. This situation is somewhat reminiscent of the apparent relative paucity of Mesozoic fossil evidence of tetrapods (mammals and birds) 58.

Several papers have noticed that molecular time estimations are consistently older than paleontological ones 2356759. Benton and Ayala 60 have pointed out four pervasive biases that make molecular dates too old: i) too old calibration dates based on previous molecular studies; ii) undetected fast-evolving genes; iii) ancestral polymorphism that is maintained through long evolutionary period; and iv) asymmetric distributions of estimated times, with a constrained younger end but an unconstrained older end (this is caused because rates of evolution are constrained to be nonnegative, but the rates are unbounded above zero).

The first factor is not the case for the present study, because we did not use the calibration dates based on previous molecular studies, but used only those based on fossil records. The third factor would be the case when the used genomic regions are under the long-term balancing selection, but no mitochondrial gene has been reported to be under such selection. Regarding the second and fourth factors, we believe that they are also not the case for this study, because we used mitogenomic sequence data excluding peculiarly rapid evolving region (e.g., the control region), and because each mitochondrial gene used here was tested to perform well for dating vertebrate (tetrapod) divergences 61. According to Benton and Ayala 60, for reliable dating "careful choice of genes may be a more appropriate strategy (than the larger data strategy), with a focus on long and fast-evolving (yet alignable) sequences." Our present study based on nearly whole mitogenomic sequence data fairly accommodates such condition.

We then conducted the divergence time estimation using the Gondwanan vicariance assumption regarding cichlids as additional time constraints (Fig. 4). Compared to the results shown in Figure 2 (without the additional time constraints), the means of estimated divergence times at various nodes are similar or somewhat larger (= 18 million years; see Table 3). However, the 95% credibility ranges of the estimated times overlap well between the two results, and the differences in mean values are not large, compared to potential error ranges in other elements, such as stochastic errors in molecular evolution and errors in dating fossils.

The addition of the cichlid constraints appears to shorten the 95% credibility intervals of the time estimates, especially for divergences occurring within Acanthomorpha 100–200 MYA. For example, our Figure 2 and Yamanoue et al. 55 estimated the divergence time of torafugu (Tetraodontiformes) and medaka (Beloniformes) to be approximately 159 (136–183) MYA and 184 (154–221) MYA, respectively. The cichlid constraints considerably narrowed the 95% credibility interval to 176 (163–191) MYA (Table 3), and also increased the precision of time estimates for other nodes. The use of ample molecular data from mitogenomic sequences also helped to narrow the credibility interval. For example, Kumazawa et al. 5 used two mitochondrial genes (NADH dehydrogenase subunit 2 and cytochrome b) and estimated the divergence between torafugu and zebrafish at 284 ± 28 (mean ± standard deviation) MYA, whereas our whole mitogenomic data set showed the divergence at 288 (268–307) MYA (Table 3).