The plots (Figures 7.1 A-D) show the faster evolution of the cytochrome b fragment compared with 16S rRNA. This is mainly caused by the high substitution rate in the third codon position of cytochrome b, while the first codon position evolves as fast as the 16S rRNA fragment (all positions including hypervariable regions), and the second evolves slower.
Figure 7.1. Plots of (A) pairwise uncorrected p-distances of 16S rRNA vs. cytochrome b; (B) pairwise uncorrected p-distances 16S vs. first position of cytochrome b; (C) pairwise uncorrected p-distances 16S vs. second position of cytochrome b; (D) pairwise uncorrected p-distances 16S vs. third position of cytochrome b. All comparisons refer to datasets 1 and 3.
The uncorrected p-distances (in percent) among mantellids for the cytochrome b gene are quite high, ranging up to 18% within the genus Aglyptodactylus, up to 30%
within the genus Boophis, up to 5% within the species Laliostoma labrosum, up to 19%
within the genus Mantella, up to 30% within the genus Mantidactylus, and ranges from 17-34% among different genera. The average uncorrected sequence divergence is 24%
among mantellids (dataset 1) and 10% among the sequences of Mantella (dataset 2) considered in this study (Table 7.1). Distances after applying the K2P correction, for congeneric individuals, reach (in percent) 38% in Aglyptodactylus, 45% in Boophis, 37% in Laliostoma labrosum, 46% in Mantidactylus and 39% in Mantella, with an average of 29% which is higher than the value obtained by Johns and Avise (1998, see Figs. 3, 5 therein) for congeneric species and confamilial genera of amphibians.
Genus Cytochrome b 16S rRNA
Aglyptodactylus 0-18% 0-10%
Boophis 0-30% 0-18%
Laliostoma 1-5% 0-0.6%
Mantella 0-19% 0-10%
Mantidactylus 0-30% 0-22%
Between genera 17-34% 10-24%
Average 24% 15%
Table 7.1. Uncorrected p-distance within each mantellid genera for the cytochrome b and the 16S rRNA genes. Because of partly unresolved taxonomy at the species level, the values summarize intraspecific and interspecific distances.
The uncorrected p-distances for the 16S rRNA gene (dataset 3) reach 10%
within the genera Aglyptodactylus and Mantella, 18% in Boophis, 0.6% in Laliostoma labrosum, 22% within the genus Mantidactylus, and range from 10-24% among different mantellid genera. The average of uncorrected sequence divergence is 15%
among mantellids (dataset 1) and 5% among the sequences of Mantella (dataset 2) compared (Table 7.1, Figure 7.2).
Figure 7.2. Box graph representing the uncorrected p-distances (in percent; y-axes) for each mantellid genus (x-axes) for the cytochrome b gene (black) and the 16S rRNA gene fragments respectively (grey).
Already the results of Graybeal (1993) indicate that plain nucleotide distances are not always a reliable indicator of the level of saturation, but the high divergences in cytochrome b sequences suggest saturation at the third codon position (Meyer 1994).
Indeed, among mantellids (dataset 1), the cytochrome b gene is highly saturated at its third, but not at the first and second codon positions (Figures 7.3 A, B, C).
Figure 7.3. Plots of pairwise sequence divergences (in percent) in the cytochrome b gene of mantellid frogs: (A) total uncorrected p-distances vs. uncorrected p-distances at first codon positions; (B) total uncorrected p-distances vs. uncorrected distances at second positions; (C) total uncorrected p-distances vs. uncorrected p-distances at third positions; (D) total uncorrected p-distances vs. uncorrected p-distances at third positions in the Mantella (dataset 2); (E) total uncorrected p-distances vs. absolute numbers of transitions (in percent) only at third positions ; F) total uncorrected p-distances vs.
absolute numbers of transversions (in percent) only at third positions. All comparisons except for D refer to dataset 1.
In contrast, analysing only values among members of the closely related species of the genus Mantella (dataset 2), cytochrome b appears not to be saturated at the third codon position (Figure 7.3 D). As expected the saturation of the third codon position is caused by high saturation of transitions, but did not yet reach saturation for transversion (Figures 7.3 E, F, but see also Table 7.3, since the skew in base composition determines the maximum sequence divergence at which saturation is reached).
The ratio between nonsynonymous and synonymous mutations (dN/dS) in pairwise comparisons among mantellids (dataset 1) ranges from 0 to 0.108, indicating that the majority of mutations are synonymous (data not shown). The highest value was observed in only one case, and the observed dN/dS ratio in general supports the assumption of neutral evolution of the cytochrome b gene although it may not be sensitive enough to detect subtle molecular adaptation (McClellan et al. 2004).
As previously observed in other vertebrates (Graybeal 1993; Farias et al. 2001;
Meyer 1993; Saccone et al. 1999; Santucci et al. 1998) there is a tendency of low content of guanine at second (G= 15%) and third position (G= 7%). First codon positions are relatively homogenous in terms of base composition, but tymine is favoured as second codon position, as already observed in birds (Friesen et al. 1993;
Groth 1998; Krajewski and King 1996). The third codon positions favour cytosine (Table 7.2). Variation among genera in degree of base compositional bias is not significant, except in Mantella and Laliostoma. The Chi-Square Test implemented in PAUP* does not reject the hypotesis of homogeneous base composition within Mantellidae.
Table 7.2. Average nucleotide frequencies in cytochrome b sequences of mantellids calculated with the program PAML.
The three samples of Laliostoma labrosum have higher content of guanine at third position than at second and at the third position there is a bias against adenine (Table 7.3). In the genus Mantella we found at the third codon position a particularly low (<2%) content of guanine and a bias favouring tymine instead of cytosine as in the
other genera (Table 7.3). The mean base composition for the cytochrome b gene is 32%
tymine, 27% cytosine, 25% adenine and 16% guanine (Table 7.2).
Codon Postion
Table 7.3. Average nucleotide frequencies in each mantellid genus for cytochrome b sequences calculated with the program PAML.
Phylogenetic Analysis
The best fitting model selected by AIC in Modeltest for the cytochrome b fragment analysed (dataset 1) was the TrN+I+G substitution model with a gamma distribution shape parameter of 0.6452 and proportion of invariable sites of 0.3242. The GTR+I+G model with a gamma distribution shape parameter of 0.6755 and proportion of invariable sites of 0.4353 was selected for the data set after exclusion of third codon positions. The best models for the data sets of only first, second, and third positions, respectively, were a TVM+I+G model with a gamma distribution shape parameter of 0.7122 and proportion of invariable sites of 0.3094, the same model with a gamma distribution shape parameter of 0.7732 and proportion of invariable sites of 0.5362, and a GTR+G model with a gamma distribution shape parameter of 2.4384 and proportion of invariable sites of 0.
Of the total of 551 characters of the cytochrome b fragment analysed (dataset 1), 184 (33%) were invariant, 39 (7%) were variable but parsimony-uninformative and 328 (60%) were variable and parsimony-informative. After exclusion of third positions, of
the total of 368 characters, 184 (50%) were constant, 39 (11%) variable but parsimony-uninformative and 145 (39%) variable and parsimony-informative.
The best fitting model selected for the 16S rRNA fragment analysed (dataset 3) was GTR+I+G with gamma shape parameter of 0.6510 and proportion of invariable sites of 0.2654. Of the total of 360 characters of the 16S rRNA fragment analysed for phylogenetic purpose, 147 (41%) were invariant, 43 (12%) were variable but parsimony-uninformative and 170 (47%) were variable and parsimony-informative.
The tree obtained by partitioned Bayesian analysis of cytochrome b (dataset 1;
Figure 7.4) agrees in many aspects with the highly supported topology of Vences et al.
(2003) that was based on a much larger data set of mitochondrial and nuclear genes (1875 bp) in a reduced but representative set of taxa (47). In contrast to that analysis, the cytochrome b data herein do not support monophyly of the subfamilies Laliostominae and Boophinae: the genus Laliostoma clusters together with the brook-breeding clade of Boophis, and the genus Aglyptodactylus is placed as sister group of some species of pond-breeding Boophis species, although without statistical support.
The subfamily Mantellidae is found to be monophyletic. Two species of the subgenus Guibemantis are not placed in the Guibemantis clade: Mantidactylus liber with Pandanusicola and M. elegans with Spinomantis. Mantidactylus bertini (subgenus Blommersia) is placed ancestral to the clade containing the subgenus Blommersia and many other mantellines, and M. argenteus with Ochthomantis. The tree from the partitioned Bayesian analysis of dataset 1 agrees in almost all aspects with the tree from the unpartitioned Bayesian analysis, but considerably differs from trees using only first and second positions, and from a tree using all codon positions under the ML and the MP optimality criteria (Figure 7.5).
Figure 7.4. Tree of 191 mantellids specimens based on Bayesian partitioned analysis.
One asterisk indicates Bayesian posterior probability 95-98%. Two asterisks indicate Bayesian posterior probability 99-100%. The absolute numbers indicate Bayesian posterior probability lower then 95%. Subgenera or genera are collapsed. The outgroup is not shown.
Figure 7.5. Collapsed phylogenetic trees of mantellids based on cytochrome b fragment: (A) ML; (B) Bayesian; (C) ML analysis after exclusion of third codon position; (D) Bayesian analysis after exclusion of third codon position; (E) MP; (F) MP analysis after exclusion of third codon position. The outgroup is not shown.
The exclusion of third codon position of the cytochrome b data set to eliminate saturation negatively affects the robustness of the tree (Figure 7.5). The exclusion of this character determines the lack of information at the genus and species-group level (shallow divergence) and creates polytomy (Figures 7.5 D, F). For example, the usually very well defined genus Mantella becomes paraphyletic (Figures 7.5 D, F), as do other well-established clades (see Discussion below). The 16S rRNA trees (ML and Bayesian) recover largely the monophyly of the subgenera and genera, although they do not support monophyly of Mantidactylus (Figures 7.6 A, B).
Figure 7.6. Collapsed phylogenetic trees of mantellids based on 16S rRNA fragment:
(A) ML; (B) Bayesian. The outgroup is not shown.
The genus Boophis that results paraphyletic based on cytochrome b is monophyletic based on the 16S rRNA fragments considered, but the relationships among different subgenera and genera in the majority of the cases do not reflect the usually well-resolved relationships (Figures 7.6 A, B). The Bayesian analysis better resolves a number of relationship as compared to ML, by (1) placing Blommersia with the Guibemantis/Pandanusicola clade instead of Laliostoma, (2) supporting monophyly of Mantella, (3) placing M. bertini with Spinomantis, and (4) placing the subgenera Gephyromantis, Laurentomantis and Phylacomantis as monophyletic sister group of the
stream-breeding Mantidactylus lineages. Altogether cytochrome b performs better in resolving a tree than the 16S rRNA fragment analysed (in dataset 3) and the latter marker performs slightly better in the Bayesian analysis than under the ML criterion (Figures 7.6 A, B).
7.5. DISCUSSION
Phylogeny of mantellid frogs
Mantellids, in the past years, have been subject to a large number of phylogenetic analyses based on morphological (Vences et al. 2002b) and molecular datasets (Chiari et al. 2004; Lehtinen et al. 2004; Lehtinen and Nussbaum 2003;
Richards et al. 2000; Vences et al. 2000a, 2002a, 2003). Very large molecular datasets of 16S rRNA sequences from over 1000 individuals have also been obtained in the context of a DNA barcoding effort of these frogs (Vences et al. 2005). Especially the highly resolved tree of Vences et al. (2003) provides a means to evaluate the accuracy of the phylogenetic reconstructions herein. In general, the various clades shown as collapsed in Figure 7.4 and Figure 7.5 have been well supported by other data sets, and the fact that most were recovered in the present analysis, even if sometimes without statistical support, gives some confidence in the utility of cytochrome b. The major disagreement with all other data sets is the non-monophyly of the Laliostominae (Laliostoma + Aglyptodactylus) and of Boophis. The corresponding nodes are not significantly supported (Figure 7.4). They are one example for the tendency of recovering wrong basal relationship using cytochrome b at deeper divergences. In general, no high support is found for any of the basal nodes, whereas significant Bayesian support values were obtained for a number of more shallow clades, such as the genera Aglyptodactylus, Laliostoma and Mantella and the subgenera Chonomantis, Hylobatrachus, Mantidactylus and Ochthomantis. An important aspect, not further tested here, might have been the dense taxon sampling in our analysis, which is known to greatly improve phylogenetic accuracy (e.g., Zwickl and Hillis 2002; but see Rokas and Carroll 2005 and the review in Cummings and Meyer submitted).
Several phylogenetic placements differ from current classification and merit discussion. The fact that the genus Mantidactylus is paraphyletic, Mantella being more closely related to some Mantidactylus than to others, is well established and needs no further discussion (Richards et al. 2000; Vences et al. 2003). Mantidactylus liber is
placed with Pandanusicola, in agreement with previous analyses based on different genes (Lehtinen et al. 2004; Lehtinen and Nussbaum 2003; Vences et al. 2003), indicating that it belongs into this subgenus despite its different morphology and reproductive biology (Glaw and Vences 1994). Mantidactylus bertini is currently classified in Blommersia, but is here placed ancestral to a clade containing this and other subgenera. This is an enigmatic species of which very little biological information has become available so far. We recently collected this species from several sites in south-eastern Madagascar and confirmed that its terrestrial calling behaviour along streams, and femoral gland morphology, differs from other Blommersia and argues for classificatory change. The high divergence between the two M. bertini sequences included here are paralleled by divergences in the 16S rRNA gene (not shown), and strongly suggest the existence of several cryptic species subsumed under the name M.
bertini. Mantidactylus madinika, originally considered to belong into Blommersia as well, is placed sister to Mantella. The same grouping was recovered by Vences et al.
(2003) and indicates that this miniaturized species may indeed be similar to the ancestor of the highly derived Malagasy poison frogs included in Mantella. A third apparently misplaced Blommersia is Mantidactylus argenteus, which here stands sister to the subgenus Ochthomantis. Also this species is morphologically and biologically different from other Blommersia, showing parental care, and recent data on parental care in Mantidactylus majori (subgenus Ochthomantis) make sense in the light of the grouping suggested here (Vences and De la Riva in press). In conclusion, several cases that appear to be weird misplacements in our tree (Figure 7.4) may indeed reflect correct phylogenetic relationships.
Saturation and phylogenetic utility of cytochrome b
In an effort to evaluate the effect of the phylogenetic signal carried by each codon position and of their noise in tree reconstruction in mantellids, we contrasted a partitioned analysis, weighing each codon position differently, against a non-partitioned analysis. Commonly, the second codon position carries more phylogenetic information than the first and the third positions. Third codon positions and transitions are often downweighted in the analysis due to their saturation (reviewed by Meyer 1994).
Saturation of cytochrome b at the third and sometimes first codon position has been
Griffiths 1997). However, third codon positions are not always misleading and their exclusion can sometimes negatively affect phylogenetic resolution (Björklund 1999;
Krajewski and King 1996; Yoder et al. 1996). Cummings et al. (1995) observed that third codon positions were phylogenetically informative if the number of sites was large and ML was used, but not with MP and NJ. An a priori evaluation of the degree of saturation is then necessary to discriminate between the information and noise carried by third codon position.
Our data showed that saturation at third codon position begins at about 13%
uncorrected total sequence divergence (Figure 7.3 C). This implies that the effect of saturation at the third codon position is high among subgenera and genera, which might lead to problems of resolution or robustness of support. Moreover, the strong codon biases with the particularly low content of guanine at the third position will cause a codon bias. This will result faster saturation on a smaller maximal sequence divergence.
The bias will also increase differences among the examined species in terms of amino acid substitutions (see also Meyer 1993, 1994). Since mutations in third codon position usually do not result in amino acid changes, they will tend to maintain a functional gene product. In mantellids, the synonymous mutations in the cytochrome b genes surpassed the amount of nonsynonymous mutations and the wide majority of nucleotide changes did not cause amino acid changes.
The a priori analysis showing the saturation at the third codon position of the used marker would suggest exclusion of this character from the phylogenetic analyses to remove the possible noise due to homoplasy. Nevertheless, the Bayesian, ML and MP analyses based on first and second positions only (Figures 7.5 D, E) are in several aspects worse than the partitioned Bayesian tree (Figure 7.4). Not considering the novel phylogenetic relationships suggested by our tree and discussed above, this can be best documented by changes in the relationship among four clades. These are well-established monophyletic groups by molecular and morphological data (e.g., Vences et al. 2003), and are recovered by the analysis of all three cytochrome b codon positions but not by the analysis after exclusion of third positions: (1) the genus Mantella; (2) the forest-dwelling frogs belonging to the subgenera Gephyromantis, Laurentomantis and Phylacomantis; (3) the subgenus Guibemantis; (4) the subgenus Brygoomantis.
Hence, mantellid relationships are better recovered including third positions, suggesting that even this highly saturated partition includes some phylogenetic
information. In addition, our data also indicate that the partitioned Bayesian analysis performed better than the unpartitioned ML and MP analyses: ML did not recover monophyly of the two specimens assigned to M. bertini and creates polytomy of the subgenus Guibemantis, while MP did not recover monophyly of the subgenera Guibemantis, Blommersia and Spinomantis. In contrast, there were no conspicuous differences between the partitioned and the unpartitioned Bayesian analysis.
The sequence divergences between genera in 16S rRNA (for the whole amplified fragment) are distinctly lower than in cytochrome b (Table 7.1). In addition, before phylogenetic analysis we have further reduced this variation, and the number of potentially informative sites, by removing gapped and hypervariable regions. This fact could explain why the cytochrome b marker better recovers the correct relationships among different genera as compared with 16S rRNA (Figures 7.4 and 7.6 A, B).
As a conclusion, the cytochrome b gene performs better than the 16S rRNA fragment analysed (after exclusion of hypervariable sites) in resolving phylogenetic relationships among mantellids. The inclusion of the 16S rRNA hypervariable sites would probably increase the phylogenetic resolution of this gene, but such a correct alignment is a highly time-consuming and error-prone task. Despite a generally good performance, the short fragment of cytochrome b used herein was unable to provide sufficient statistical support for many relationships and therefore requires combination with other markers. Our data suggest that at the level of species, subgenera and genera of frogs, third positions of this gene do bear phylogenetic information and therefore should not be a priori excluded from analysis. Partitioned Bayesian analysis appears to outperform ML and MP for such datasets of short sequences of limited phylogenetic signal.
ACKNOWLEDGMENTS
We are indebted to Dirk Steinke for his comments on this manuscript and for helping with the TaxI program. We are also grateful to Arie van der Meijden and Simone Hoegg for valuable comments on the first draft of the manuscript, to Walter Salzburger for helping with the PAML program, and to Franco Andreone, Frank Glaw and David R.
Vieites for field companionship during sample collection. The Malagasy authorities kindly issued research and export permits. Y. Chiari was supported by an Ambassadorial Scholarship of the International Rotary Foundation. Laboratory work received support from grants of the Deutsche Forschungsgemeinschaft to M. Vences and A. Meyer and the University of Konstanz to A. Meyer.