© 2005 Society of Systematic Biologists
Exploring Rate Variation Among and Within Sites in a Densely Sampled Tree: Species Level Phylogenetics of North American Tiger Beetles (Genus Cicindela)
Edited by Ziheng Yang: Assiciate Editor
1 Department of Entomology, The Natural History Museum London SW7 5BD, United Kingdom E-mail: a.vogler{at}nhm.ac.uk
2 Department of Biological Sciences, Imperial College London Silwood Park Campus, Ascot, Berkshire SL5 7PY, United Kingdom
3 Faculdade de Ciências da Universidade de Lisboa, Departamento de Biologia Animal, Centro de Biologia Ambiental, Rua Ernesto Vasconcelos 1749–016, Campo Grande, Lisboa, Portugal
4 Jodrell Laboratory, Royal Botanic Gardens Kew TW9 3DS, United Kingdom
 |
Abstract |
|---|
 Top Abstract Material and MethodsResultsDiscussionConclusionsAppendix 2AcknowledgmentsReferences |
|---|
Species-level phylogenetic studies require fast-evolving nucleotide positions to resolve relationships among close relatives, but these sites may be highly homoplastic and perhaps uninformative or even misleading deeper in the tree. Here we describe a species-level analysis of tiger beetles in the genus Cicindela (Coleoptera: Cicindelidae) for 132 terminal taxa and 1897 nucleotide positions from three regions of mtDNA, comprising 75% coverage of species occurring in North America. Evenly weighted parsimony analysis recovered four major clades representing radiations confined to North and Central America. Relationships near the tips were well supported but signal was contradictory at deeper nodes. Two major categories (3rd positions and all others) can be distinguished in likelihood analysis of character variation, of which only the fast-changing 3rd position characters were affected by saturation. However, their downweighting under a variety of criteria did not improve the tree topology at basal nodes. There was weak conflict between 3rd and non-3rd position characters deep in the tree, but support levels declined towards the root for all categories, even on trees that were reconstructed from 3rd and non-3rd positions separately. Statistical analysis of parsimony-based character transitions along branches showed a largely homogeneous distribution of change along the root-to-tip axis. The comparison of character transitions among the four major portions of the tree revealed deviations from stochastic distribution for the non-3rd positions, but not for 3rd positions. Hence, variability of functionally constrained non-3rd positions differs between clades and may be dependent on the character states at other sites, consistent with the covarion model of molecular evolution. The results suggest that some properties of 3rd positions are less problematic for phylogenetic reconstruction than other categories despite their high total homoplasy. In densely sampled data sets of closely related species, the disadvantages of weighting schemes according to homoplasy levels outweigh the benefits, showing the difficulty of devising meaningful weighting schemes that are applicable universally throughout the tree.
Keywords: Coleoptera; heterotachy; homoplasy weighting; mtDNA; nodal support; rate variation; saturation; species level phylogenies
Received November 1, 2001; Revised April 5, 2002; Accepted August 17, 2004
Within the Coleoptera (beetles), there are some 20 genera with 1000 or more species. Among these, tiger beetles in the genus Cicindela are perhaps the best known with respect to taxonomy, ecology, and geographic distribution and therefore represent an ideal group for investigating the evolutionary causes of extreme species diversity (Pearson and Vogler, 2001). Phylogenetic trees are vital for such studies, to analyze the evolution of characters thought to have promoted species richness, and sequence variation can be used to estimate the time of diversification. In particular, trees that include all extant taxa in a clade ("species-level phylogenies") are used for phylogenetic analysis of character changes which may be instrumental in speciation (Barraclough and Nee, 2001). As a basis for such studies we present a phylogenetic analysis of the North American Cicindela based on mtDNA, to include around 75% of all species occurring in this biogeographical region.
An important issue for building species-level phylogenetic trees is the choice of markers and their treatment in the analysis for phylogenetic reconstruction at various hierarchical levels. Whereas the separation of close relatives requires the use of fast-evolving sites, these may be highly homoplastic throughout the wider tree and uninformative or even contradictory regarding deeper relationships. It has been suggested that homoplastic variation should be assigned lower significance during tree reconstruction, by downweighting during tree search (e.g., Farris, 1969; Goloboff, 1993; Cunningham, 1997). However, if fast-evolving sites are responsible for supporting relationships among the closest relatives, this strategy will lower the confidence in the tree at the species level, where the character variation may be perfectly consistent with the tree (Börklund, 1999; Kallersjo et al., 1999).
If homoplastic character variation can provide a phylogenetic signal that is consistent with the tree locally, this requires that rates of change in fast-evolving sites are not clustered in the tree; i.e., that characters evolve at similar rates in different subclades and throughout a lineage from root to tips. According to the covarion model of evolution, a certain nucleotide position is subject to change in some lineages but not others, and membership in the variable or invariable category may change over time (Fitch, 1971; Miyamoto and Fitch, 1995; Lockhart et al., 2000), leading to shifts in evolutionary rate for a given character throughout the tree (heterotachy) (Philippe et al., 2003). These rate shifts at given sites may lead to homoplasy being accumulated in a particular portion of the tree, further reducing the information content of affected characters. Rate changes are expected to impact tree reconstruction and decisions about weighting regimes, and also could affect clock estimates in various parts of the tree.
To test these issues about phylogenetic information content of homoplastic characters and the use of weighting in phylogenetic reconstruction, we attempted to dissect character variation and its distribution throughout the tree. We tested the effect of removing or downweighting the most homoplastic classes of characters on tree topology and support levels and the possibility of establishing better supported topologies on the basis of less homoplastic variation, in particular at the deeper nodes. The findings provide a clearer picture of rate heterogeneity within and among character partitions and across the tree (among subclades and along the root-to-tip axis) and help to assess their effect on tree reconstruction and support in phylogenetic analyses at the species level.
|
Material and Methods |
|---|
|
TopAbstractMaterial and MethodsResultsDiscussionConclusionsAppendix 2AcknowledgmentsReferences |
|---|
Laboratory Procedures and Phylogenetic Analysis
The North American fauna of Cicindela includes 147 described species, of which 98 occur north of Mexico (Boyd, 1982; Pearson et al., 1997). An extensive taxonomic treatment of the New World fauna was provided by Rivalier (1954), who investigated them as part of a larger taxonomic study of Cicindela (sensu lato) worldwide (Rivalier, 1950, 1954, 1963). Some 12 of the recognized subgenera are found in North America. We aimed to obtain samples for as many species from Canada, the U.S., and Mexico as possible, to present a species-level analysis of the North American fauna. Our final sample of 132 terminal taxa includes: 83 species found in America north of Mexico (Pearson et al., 1997), 24 species found in Mexico but not further north, 9 subspecies or putative new species from the study region, 9 Central American or South American species, 1 European species, and six outgroups from other genera of tiger beetles. The coverage is 85% of the species north of Mexico (Pearson et al., 1997), and 49% of species found in Mexico but not further north (Boyd, 1982). Data reported here are new for 78 taxa, with the remainder already published previously (Vogler and Welsh, 1997; Vogler and Kelley, 1998; Vogler et al., 1998; Barraclough et al., 1999) (Appendix 1). Missing species were those for which we were unable to obtain samples, many of them rare or endangered.
All procedures for DNA extraction and polymerase chain reaction (PCR) amplification were described previously (Vogler and Welsh, 1997). The aligned matrix contains 1897 base pairs (bp), comprising 411 bp of cytochrome b (CytB), 650 bp of cytochrome oxidase III plus tRNAGly (COIII), and 836 bp of 16S rRNA and the adjacent tRNALeu and ND1 genes (16S). The data matrix was largely complete, except for four species each without data for 16S and CytB, and 11 species without COIII data (Table 1). As is typical for mtDNA in insects, all regions exhibited high A-T bias (69.1% A-T for the entire data set, 72.7%, 65.8%, and 66.6% for the 16S, CytB, and COIII regions, respectively). Alignment in the ingroup was trivial because no length variation was observed, except for a single base pair in 16S rRNA of C. lemniscata. Outgroups differed in length by up to 10 base pairs and were aligned manually.
|
Tree searches were conducted with PAUP4.0b10. Parsimony trees were obtained after 1000 random addition searches with TBR branch swapping, keeping one tree per replicate. A subsequent search on these shortest trees was conducted with the Multrees option. Bremer support (BS) and partitioned Bremer support (PBS) values were determined after producing constraint files with TreeRot (Sorenson, 1999). Finding optimal trees as the prerequisite for calculating the correct BS values is problematic due to the size of the data set. Two search strategies were applied to obtain shortest trees: (1) by performing 300 random addition replicates keeping only a single tree each, followed by a second search on the shortest tree(s) from the first round keeping up to 200 trees (maxtrees = 200); and (2) performing 100 random addition replicates and collecting no more than 50 trees from each replicate (multrees = yes, nchuck = 50, chuckscore = 1). On several occasions PBS values differed substantially between trees obtained from either search strategy, even if the BS values were identical. In these cases, the average of the PBS scores was used for the analyses.
To examine the contribution of different data partitions, we also searched for shortest trees using subsets of the data—each of the three major DNA regions in turn, and each functional partition (RNA, 1st, 2nd and 3rd codon positions)—separately. For each analysis we counted the number of nodes in common with the analysis from the entire data set and calculated the increase in tree length obtained when the entire data were fitted to the tree topologies obtained from each partition. Tests for incongruence among partitions were performed using the ILD test (Farris et al., 1994) as implemented in PAUP. Despite recent criticism of this test (Dolphin et al., 2000; Yoder et al., 2001; Barker and Lutzoni, 2002; Darlu and Lecointre, 2002), it was used here as a useful heuristic tool in assessing the effects of combined and separate analysis of data partitions on tree length. We also used a measure of incongruence based on PBS values, testing for the Spearman rank correlation of PBS values at given nodes for pairs of partitions according to Sota and Vogler (2001). Heuristic methods for assessing character support, such as BS, remain valuable. Statistical methods including posterior probabilities in Bayesian analysis have expected behaviors and a more direct correlation with the level of support, but uncertainty remains because the levels of robustness in these calculations rely on the assumption that the model is right.
ML branch lengths, used for various calculations of rates of change, were calculated in PAUP based on the topology of parsimony trees. To assess saturation of character change, uncorrected parsimony branch lengths were plotted against branch lengths reconstructed by maximum likelihood (ML), assuming a GTR+
+I model. ML presumably provides a more accurate measure of branch lengths at high levels of substitutions which is underestimated in parsimony reconstruction for highly variable sites. Hence the plot is expected to level off at longer branch lengths when there are multiple changes along branches. The analysis of saturation was conducted for the entire data set and each major functional partition separately. To assess major rate categories, we compared simple GTR and GTR+site-specific (GTR+SS) rate models.
Distribution of Character Changes across the Tree
The uniformity of character variation across the tree, i.e., whether the relative rates at sites are consistent among different parts of the tree, was tested as follows. The number of steps for each character was printed in MacClade version 4.0 using the Character List command. These data were saved as a table and imported into an MS Excel spreadsheet for statistical analysis. We tested whether the relative rate of change of each character tends to be the same across subclades recognized from the parsimony analysis using a chi-square contingency table. The number of steps of each character reconstructed by MP were recorded separately for each subclade. Hence, each entry in the table is the number of steps of character r in subclade c, where r and c refer to row and column numbers, respectively (only variable characters were included in the table). The expected number of steps of a site r in a clade c, assuming that the probability of changes at a site is the same among subclades, is:
|
|
We calculated the chi-square value for the entire table (Sokal and Rohlf, 1995), with degrees of freedom equal to (R – 1) x (C – 1), where R is the number of variable sites and C is the number of subclades. A significant value would indicate that given characters tend to be more variable in some subclades than others. A nonsignificant value would indicate that characters tend to evolve at the same rate in all subclades tested. Because expected frequencies of many cells were below one, the observed chi-square value will be biased (Sokal and Rohlf, 1995). Therefore, instead of using look-up tables, we assessed the significance of the observed value by Monte Carlo simulation. Holding the total number of steps at each site constant, we assigned changes randomly to each of the subclades in proportion to the total number of steps in that clade and repeated the procedure 1000 times to obtain the expected distribution of chi-square under the null hypothesis.
We also tested whether the relative rates of change of each character are homogeneous with respect to the level of each branch from the root and the tips. As a measure of the level of each branch we used the height of its ancestral node when the tree is drawn as a rectangular cladogram; i.e., the terminal branches of a pair of sister species are level 1, the branch connecting them is level 2, etc. Then, for each site we added up the total number of changes occurring on branches of each of four categories: level 1, level 2 to level 5, level 6 to level 10, and above level 10. We performed a chi-square test on the resulting contingency table, with rows representing sites and columns representing the branch level category. A significant result would indicate that sites tend to differ in their relative rates of change between branch level categories.
To visualize the data further, we recorded which sites were reconstructed to change on each branch of the tree. We measured the average variability of the sites for which a change is inferred on a given branch and then calculated the average of the number of steps of those sites on that branch. This value was calculated for each branch of the tree. We then plotted this average variability value obtained for each branch against branch level. Under the scenario that the characters changing on a particular branch are sampled from the overall set of characters in proportion to their variability across the entire tree, we expect no relationship between average variability of characters on a branch and branch level.
|
Results |
|---|
|
TopAbstractMaterial and MethodsResultsDiscussionConclusionsAppendix 2AcknowledgmentsReferences |
|---|
Parsimony Analysis and Phylogeny of North American Cicindela
Tree searches on the entire character matrix resulted in 12 shortest trees of 8744 steps. The level of nodal support varied but was generally low for the backbone of the tree whereas relationships of many smaller groups near the tips were strongly supported (Fig. 1). We defined four major clades which conform largely to the traditional classification of Rivalier (1963) and findings from earlier analysis based on a subset of the data (Vogler and Kelley, 1998): (1) Clade 1, representing Rivalier's subgenus Cicindela (s.str.); within this group two major clades can be recognized corresponding to the C. maritima species group (group IV of Rivalier), and a clade of all other Cicindela (Rivalier's groups I, VII, VIII, plus minor subgenera Pachydela and Tribonia and two species of Cicindelidia, Ci. willistoni, and Ci. armagosae). (2) Clade 2, corresponding to Cicindelidia, minus Ci. willistoni, Ci. armagosae, and Ci. trifasciata (included in clade 4); within Cicindelidia, species exhibiting an aposematic "orange abdomen" (Pearson et al., 1988) were monophyletic and the "black abdomen" group and the aurora group can be separated from these. (3) Clade 3, comprising Ellipsoptera, Dromochorus, and a subset of species within Rivalier's (1963) Cylindera. (4) Clade 4, centered around Habroscelimorpha and including Rivalier's subgenera Microthylax, Opilidia, and Eunota; species in this group are associated with ocean beaches and salt flat habitats. With the exception of clade 1, which is likely to include several Palearctic species, each of these four clades constituted an independent endemic radiation within North America. In addition to these four groupings, the North American fauna comprised representatives of groups with mostly Neotropical distribution, in particular from Rivalier's subgenera Cylindera, Plectographa, and Brasiella, which are placed near the base of the tree. The specific relationships and the implications for classification of Cicindela will be discussed elsewhere (Duran and Vogler, 2004).
|
Tree Statistics and Level of Congruence in Data Partitions
The tree was based on three protein-coding genes as well as structural RNAs, from three separate gene regions of the mtDNA. Each gene region represented roughly similar numbers of potentially informative sites, but the total number of base pairs sequenced was higher for the 16S region than the other two regions and included a higher proportion of invariant sites (54.9% versus 47.5% in the COIII and 47.4% in CytB regions; Table 1). The number of steps in parsimony trees was lowest in 16S, which also exhibited the highest internal consistency of all three regions (CI of 0.300 versus 0.163 in COIII and 0.130 in CytB; Table 1). In the separate analysis of the three mtDNA regions, most of the deeper nodes were not supported (not shown). Only about half of the nodes of the simultaneous analysis tree were also recovered in the analysis of single data partitions, but this number was low in part due to the many unresolved nodes in the separate analyses rather than due to incongruence of data partitions (Table 1).
Similarly, when the data were partitioned according to functional criteria, separated by codon position or gene function (protein coding versus structural RNA), trees greatly differed in topology and resolution. With 5892 steps, the 3rd positions produced by far the greatest number of character changes in separate analysis, compared to a shortest tree of 2764 steps from all remaining characters, and 1428 steps of the shortest tree from the structural RNA. The 3rd position also recovered the greatest number of nodes from the simultaneous analysis, recovering 86 of 129 nodes obtained with all data, but only 65 nodes in trees from non-3rd and 68 nodes from structural RNAs (Table 1). However, these differences of separate and simultaneous analyses were not reflected in very strong character conflict: when the entire data set was fitted to the tree topologies obtained from individual partitions, the increase in tree length was only between 2% and 3%. The smallest partition, 2nd positions, was an exception as it produced a tree of only 209 steps, which was highly unresolved and much less congruent with the total data (constrained topology is 22.0% longer than the shortest tree; Table 1).
Character incongruence as determined by ILD was low, with simultaneous analysis adding no more than 1% to 3% to the number of steps over the separate analysis. The ILD (added numbers of parsimony tree length in combined analysis, normalized for the total number of steps in the combined analysis tree) was generally higher whenever comparisons are made between structural RNA and the other partitions (Table 2). However, this was not reflected in the Partition Homogeneity Test (Farris et al., 1994), which produced significant P values only for the comparison of 2nd and 3rd positions. The latter would be unexpected, given the presumably similar phylogenetic history of nucleotides within codons, but a rejection of the hypothesis of congruence in this case may be due to artifacts, as the significance in the ILD test may be high simply due to the difference in the levels of sequence variation in these two partitions (see below) rather than true incongruence (Dolphin et al., 2000). We also included a further test which determines congruence for a pair of data partitions based on the Spearman rank correlation on PBS values (Sota and Vogler, 2001). This analysis provides a measure of incongruence that takes into account tree topology and the magnitude of the support. This test produced significantly positive correlation for 2nd positions and structural RNA, but 2nd and 3rd positions were negatively correlated, consistent with their high ILD. The implied incongruence in this case may be due to the very different tree topologies obtained with the much smaller number of variable character in the 2nd positions, which may mislead this analysis. When the data were partitioned by gene regions, the CytB partition was significantly incongruent with the other two regions in the ILD test but not in the Spearman rank correlation, whereas COIII and 16S did not exhibit significant conflict in either test. In conclusion, although some conflict among data partitions was apparent, this might be due largely to the different size of partitions and differences in rates, rather than phylogenetically divergent signal.
|
Rate Variation among Partitions and Saturation
Rates of change differed greatly between partitions. The 3rd codon positions represented by far the most variable set of characters, with only 7 of 377 positions invariable and 4 positions apomorphic, compared with 408 of 758 sites invariable and 135 apomorphic in the structural RNA partition. Both partitions exhibited among-site rate variation, as the variation in numbers of steps among sites was significantly greater than expected assuming a Poisson distribution with the observed average number of steps per site (in both cases P << 0.0001, based on a chi-square test for goodness of fit; Sokal and Rohlf, 1995).
We conducted tests of saturation of character change by plotting parsimony branch lengths for various partitions of the data against each other. This analysis was conducted on the ingroup taxa only, as outgroup taxa were highly divergent. When outgroup taxa were excluded, parsimony searches produced 12 shortest trees of 7709 steps that were identical to the ingroup topology (Fig. 2). Plots of branch length indicated saturation of transitions relative to transversions, in particular when only 3rd-position characters are considered, but not for structural RNAs or 2nd positions (Fig. 3). The plots for 3rd positions against non-3rd positions leveled off, in particular when plotted against the more slowly evolving transversions, indicating that for the longer branches the 3rd positions experience multiple changes more frequently than the other types of characters. There was no evidence of saturation among 2nd positions or structural RNA sites (non-3rd positions) according to this test. Similar conclusions were obtained from plotting uncorrected parsimony branch lengths against branch lengths estimated by ML. Whereas the ML and parsimony branch lengths plots were closely correlated for structural RNA and non-3rd positions, plots for all sites and for 3rd positions showed greater sequence divergences in the likelihood estimate (Fig. 3), indicating that multiple changes occurred on the long branches not recovered in the parsimony reconstruction.
|
|
Additional evidence for the fundamental differences between 3rd and non-3rd characters came from further maximum likelihood optimizations. We compared the fit of a combined analysis under nested GTR and GTR+SS rate models. The likelihood under the GTR model (–ln = 47,774.49) was significantly worse than under a site-specific model which specifies two partitions, 3rd and non-3rd positions (–ln = 41,937.01;
= 5837.48). Under partitioning according to other criteria, e.g., the three gene regions corresponding to CytB, COI, and 16S, this effect of a site-specific model was much smaller (–ln = 47,118.74; = 655.75). This result confirmed that the major recognizable differences in the substitutional processes were between fast-evolving 3rd positions and more slowly evolving non-3rd positions, including 1st and 2nd positions plus those coding for structural RNAs. However, when a gamma-distributed rate heterogeneity was incorporated into the model, this could accommodate both partitions (–ln = 39,251.82).
Distribution of Character Changes Across the Tree
Rates of character change were studied in various partitions across different clades, and along the root-to-tip axis, to test for uniformity of variation across the tree. First, we assessed bias in nucleotide frequencies using a simple chi-square test of homogeneity implemented in PAUP. This test revealed an overall highly significant deviation from stochastic nucleotide composition for the entire data set and a marginally significant deviation when each of the four major North American clades were analyzed separately. However, bias in base composition was entirely attributable to the 3rd-position characters, whereas non-3rd positions were distributed evenly according to this test (data not shown).
Next, we analyzed character variation along the branches in comparison of the four major lineages established above (clades 1 to 4). Because the taxon sampling within these clades represented a largely complete set of species, rates of change per internode can be compared directly among these clades. We found the rates of change per branch to be rather uniform among different parts of the tree with the exception of clade 4, which exhibited approximately twofold higher rates of change per branch (Table 3). Rates of change differed greatly between various data partitions (structural RNA and 3rd positions, and within the 3rd positions from three different gene regions), but they were maintained at similar ratios across all four clades, including the fast-changing clade 4, suggesting that rate variation was correlated throughout the mitochondrial genome. This analysis demonstrated a similar amount of inferred change, on average, apportioned to each branch in the major subclades of the tree, with the exception of clade 4. Hence, phylogenetic reconstruction should not greatly be affected by different average branch lengths in the main clades. Higher rates per branch in clade 4 could be affected by higher rates of cladogenesis or higher rates of character evolution, but because the greater total branch lengths from the tips to the root was higher in this lineage (Table 3), faster rates of evolution were implicated.
|
Further, we assessed whether the character variation across different clades can be attributed to a largely homogeneous variation at specific nucleotide sites. However, the chi-square analysis for consistency of character variability among the clades indicated a more complex situation. Across all sites, the observed number of steps in each character in each clade differed significantly from the expected number of steps under the null model that each character evolves at the same rate in each clade (chi-square = 2557.9, P from Monte Carlo test = 0.01). However, when analyzed separately, 3rd positions showed no significant departure from even changes among clades (chi-square = 1049.1, P from Monte Carlo test = 0.91), whereas non-3rd positions showed highly significant departure from even rates of changes (chi-square = 1508.7, P from Monte Carlo test < 0.001).
Finally, we tested for uniformity of rate variation at sites along the root-to-tip axis. Node levels were categorized in four levels (levels 0, 1 to 4, 5 to 9, and above 10). Sites with a larger number of changes overall also tended to be those with a larger proportion of changes at each of the four node-level categories used in the contingency table analysis (all P-value from the Monte Carlo simulations were greater than 0.4, irrespective of whether all sites, 3rd positions, or non-3rd positions were included in the analysis). However, when the average variability of the characters changing on a branch was plotted against node level (Fig. 4), two weak departures from an even distribution with node level were observed. First, there was a slight decline in the average variability of characters towards the root, consistent with saturation of faster-evolving sites. Second, there was a counterintuitive excess of changes in slow characters on some branches at the tips (Fig. 4). One possible explanation for the latter trend could be the effects of sequence error or noise (see Discussion).
|
Effects of Character Weighting on Tree Topology and Nodal Support
Given these differences in rates, homoplasy was not uniformly distributed and hence different weighting schemes would be expected to affect various branches of the tree differently. Therefore, we tested the effects of character weighting on tree topology, using a tree from ingroup taxa only. Tree topologies were compared with the equal weight tree establishing the number of nodes in common and the recovery of major clades (Table 4). Varying the weighting of the more abundant transitions relative to transversions had little effect on the tree topology over the range of weights evaluated here (tv/ti = 2 and 3), in particular with regard to the monophyly of subgenera and other major groups. However, removal or substantial downweighting of 3rd positions resulted in greatly different topologies and reduced levels of resolution. Implied weighting according to Goloboff (1993) under a variety of parameters of the weighting function, k, also resulted in substantially different trees, in particular under settings with homoplastic changes most strongly downweighted (k = 1) (Table 4).
|
Despite a large number of nodes not found in a strict consensus with the equal weight tree, relationships of major groups generally held up between weighting schemes. Differences in tree topology frequently affected the placement of a few species whose positions were unstable. Among others, this includes: the placement of H. auraria and Ci. trifasciata which exhibited affinities to clade 3, but frequently grouped with portions of Cicindelidia (clade 2); the Ci. aurora group which was frequently found polyphyletic and not always associated with the remaining Cicindelidia; the monophyly, or otherwise, of Microthylax and its position within clade 4; the relationships within the Cicindela s. str. (clade 1), in particular the position of C. armagosae which varied widely; and the basal relationships in Cicindelas. l., including the sister relationship of Cicindela s.str. and Cicindelidia (Table 4). Table 4 also compares the effect of reducing the data set to one of the three mtDNA gene regions, which invariably resulted in great changes in tree topology. The findings may indicate that a sufficiently large data set was required for tree recovery, and this could neither be provided by a single mtDNA region nor by an analysis with 3rd positions heavily downweighted or removed entirely.
We also investigated support levels, to test whether the 3rd positions provide contradictory signals at deeper nodes as predicted. Nodal support values from the equally weighted analysis of all data (see Appendix 3 [available at the Society of Systematic Biologists website, http://systematicbiology.org] for BS values for each node) showed a clear decline at deeper nodes (Fig. 5A). When the BS values were further dissected to test for the contribution of 3rd and non-3rd positions separately, the analysis of PBS revealed that nodes from levels 1 to 5 tend to be supported predominantly by faster-evolving 3rd positions, but with non-3rd positions also providing notable positive support (Fig. 5A). From levels 6 to 10, non-3rd positions provide the majority of support, but 3rd positions still provide, on average, weak positive support for those nodes. However, for node levels 11 and above, strong support from non-3rd positions is counteracted by negative support from the 3rd positions. This might indicate that 3rd positions were weakening support at deep levels and perhaps caused the overall decline of support levels towards deeper nodes. However, when BS values were plotted against node height for trees obtained from non-3rd positions alone (Fig. 5B), this plot, too, showed the decrease of support values from the tip to the base of the tree and perhaps indicates a general trend in the data of lower BS at deeper nodes. In summary, the analysis demonstrated that downweighting of 3rd positions might reduce conflict at the tree base somewhat. But because neither partition provided strong support for deep nodes, and because there were fewer nodes deeper in the tree (only 8 nodes above level 11, but 95 nodes from levels 1 to 5), the costs of downweighting would greatly exceed any possible benefits.
|
Hence, 3rd positions appear to be informative despite their high level of homoplasy. This was also supported by a comparison with our earlier studies based on a much less densely sampled tree (Vogler and Welsh, 1997; Vogler and Kelley, 1998). The inclusion of many additional ingroup species in the current study resulted in a drop of CI due to the increasing number of changes in a given character across the entire tree, but caused a clear increase of the RI (Table 5). This indicates that the fraction of the variation that is synapomorphic increased dramatically when additional branches were included in the tree (Kallersjo et al., 1999). Downweighting or removal of the most homoplastic characters would eliminate data that are phylogenetically informative in some portion of the tree.
|
|
Discussion |
|---|
|
TopAbstractMaterial and MethodsResultsDiscussionConclusionsAppendix 2AcknowledgmentsReferences |
|---|
Patterns of Character Change, Saturation, and Downweighting
The study provided a solid phylogenetic hypothesis for the North American species of Cicindela that was in good agreement with the traditional classification (Rivalier, 1954) and previous conclusions about the evolution of ecological traits (Vogler and Goldstein, 1997). The three regions of mtDNA produced a broadly congruent signal. However, patterns of character variation were complex, where principally 3rd and non-3rd positions showed differences with regard to overall rates, nucleotide bias, among-site rate variation, and variation of rates at single sites. We were specifically interested in the question about homogeneity of character variation across the four main clades, to assess the phylogenetic information content of highly variable mtDNA characters that may be reduced if homoplasy is clustered in certain portions of the tree. The comparisons of clades 1 to 4 revealed substantial rate heterogeneity among data partitions (1st, 2nd, and 3rd codon positions and structural RNA; different mtDNA gene regions), but these differences in rates among partitions were very similar in the four clades. Hence, among-site rate heterogeneity attributable to various functional partitions was high, but it was largely uniform in different portions of the tree. Unexpectedly, when rate heterogeneity was tested for individual sites, rather than functional categories, rates were not uniform in the non-3rd positions among the four clades; i.e., there is within-site rate variation in these characters. This finding is consistent with the covarion hypothesis (Fitch, 1971), which proposes that sites differ in their propensity for variation in the context of covarying sites. Our analysis revealed these effects at the DNA level and within a fairly narrow clade.
This kind of lineage-specific rate variation suggests that the affected characters are under functional constraints (in the context of other changes). These constraints also are likely to determine the very different relative rates between functional partitions observed in non-3rd positions. In contrast, the 3rd positions did not show this kind of within-site rate heterogeneity, they showed much greater overall rates of change, and they exhibited nucleotide compositional bias not apparent in the non-3rd positions. Their pattern of variation indicates fewer functional constraints in the 3rd positions, including lineage specific effects. Instead, they may be affected by stochastic effects or selection on base composition which differs among portions of the tree.
This analysis has implications for phylogeny reconstruction, as rate variation has frequently been proposed to be taken into account in character weighting during tree search. The commonly applied approach of using levels of character change as a criterion to either downweight or remove homoplastic characters presumes that data can be partitioned into categories of fast and slowly changing characters; that faster-changing characters have less phylogenetic information content; and that the negative effects of homoplasy, if they exist, are universal throughout the entire tree.
From our analysis it clear that these assumptions may not be fulfilled. Although we were able to identify two major rate categories corresponding to 3rd and non-3rd base positions, there remained additional rate variation within each of these partitions that did not correlate with any obvious functional partition. Neither the 3rd nor the non-3rd position changes were Poisson distributed, and absolute rates within the non-3rd positions differed greatly at sites. This would prevent the application of universal weighting schemes, although the situation is perhaps less complicated due to the fact that these rate differences were fairly evenly distributed across the entire tree (at various hierarchical levels and among subclades). However, the within-site rate variation observed in non-3rd positions cannot be taken into account in these weighing schemes, highlighting the difficulty of applying universal weights to a complex data set. The limited success of weighting schemes was also apparent from various weighted tree searches. Over a wide range of weighting parameters, the differences among trees were comparatively minor and they mostly affected nodes that were poorly supported in the equally weighted analysis, indicating that major features of the tree are recovered correctly despite the high levels of homoplasy. This analysis therefore does not support the notion (e.g., Goloboff, 1993; Swofford et al., 1996; Cunningham, 1997; and many others) that the downweighting of fast evolving data partitions improves the recovery of phylogenetic signal, if patterns of among-site rate variation are heterogenous among clades.
A related issue is to establish that a particular character set is "saturated" and hence should be subject to downweighting. Procedures for establishing saturation are not straightforward. The frequently used plots of pairwise sequence divergences as a test for the level of unobserved multiple character changes may be inappropriate (see Yang, 1998). Assessing character variation on the branches of the tree (Fig. 3) might circumvent this problem, but the question remains to what extent multiple hits would actually mislead tree inference and support levels. As has been shown in simulation studies, high substitution rates per se would not result in poor performance in parsimony reconstruction, whereas rate heterogeneity would (Yang, 1998). Hence, although affected by saturation, in some respects the 3rd positions fulfill criteria of phylogenetically informative characters better than the other categories due to their greater homogeneity of substitution rates within and among sites and among clades. Their apparent heterogeneity in nucleotide composition, although highly significant, may not greatly confound the phylogenetic signal, as a LogDet distance analysis (with proportion of invariant sites estimated to 0.5033) did not result in a greatly different tree, at least at the higher node levels within the four clades (Appendix 2).
Effect of Homoplasy and Nodal Support
Based on PBS values, nodal support generally declined at deeper nodes, and the tree topology at deeper nodes was determined mostly by non-3rd positions and contradicted by 3rd positions; i.e., alternative topologies exist that were shorter for 3rd positions (Fig. 4). Therefore, the analysis of nodal support might confirm the widely held perception that 3rd-position characters are misleading and lack information, perhaps due to their high rate of variation, and hence supporting arguments for their downweighting. However, when support values were calculated for trees obtained from each partition separately, we found that BS values still declined towards the root for non-3rd positions. This indicates that the declining support is a general phenomenon of the data rather than primarily a result of conflict between fast and slowly evolving characters (which potentially could be corrected by downweighting).
In addition, the notion of downweighting to minimize conflict between signal from fast and slowly evolving characters is questionable on theoretical grounds. Third positions at deeper nodes may still contribute to the phylogenetic signal even if their PBS score is low, as they may lend "hidden support" (Gatesy et al., 1999) which only emerges in combination with other types of data. Although mostly investigated in studies where partitioning was by different genetic loci, hidden support should also apply within a single locus when rates and types of character changes within a gene partition are inhomogeneous. Hence, data interaction may be an important issue for recovery of the signal, and downweighting of faster-evolving characters would be reducing the strength of this interaction.
But why does support decline towards the root of the tree, even when downweighting schemes are applied? One possibility would be that the characters assigned highest weight in the weighting scheme are also saturated at deeper levels in the tree; i.e., that the entire data set is saturated at the deepest nodes. This does not appear to be the case for our data, because plots of uncorrected versus corrected branch lengths revealed no evidence of saturation in 2nd position or RNA sites. A second possibility is that the branching events represented by the deeper nodes were in reality close together in time, making it hard to resolve the exact order of branching (Walsh et al., 1999). We have no independent evidence to assess this possibility for our data, but it must be kept in mind. Third, the poorly supported nodes are those defining relationships among the major clades of Cicindela in North America, and, therefore, resolving the true branching patterns at those levels would need a much more representative sample of taxa from the rest of the world. It seems likely that the major North American clades are not each others' closest relatives, but have been derived independently by subsequent invasions into the continent. Finally, the reduction in support levels towards the basal nodes appears to represent a more general phenomenon of parsimony trees or species-level trees in particular, and has been observed also, e.g., in a fully sampled tree of Ips bark beetles (Cognato and Vogler, 2001). Rather than biological processes to cause these patterns, the possibility exists that some unknown effect of the reconstruction procedure is responsible (e.g., as in Collins et al., 1994), in particular due to the uncertainties of reconstructing character transformations in highly variable characters throughout a large tree. The significance of BS values, and the PBS in particular, will require further study. As these values are based on comparisons of topologies, rather than statistical properties of the data, the direct comparability of these values is compromised, and the values constitute a complex phenomenon of character distributions and branch lengths and only an indirect measure of conflict between partitions, as has been pointed out by previous authors (e.g., DeBry, 2001). Assessing PBS values at different node levels in simulated data sets may be useful to test these issues.
Although we conclude that homoplasy weighting is inappropriate for the current data set, and similar conclusions have been reached in species level studies elsewhere (e.g., Baker et al., 2001; Cognato and Vogler, 2001), other studies had good success with implementing complex weighting schemes (e.g. Cunningham, 1997; Mitchell et al., 2000). The main difference between those studies and ours is in the level of taxon sampling and overall scope. We attempted to establish relationships among taxa probably separated from one another within the last few million years (Freitag, 1965; Willis, 1967; Barraclough and Vogler, 2002), and taxon sampling is very complete at the species level. The distribution of branches will be different to that found in phylogenetic studies aiming to reconstruct higher-level relationships among more distantly related groups, in which exemplar taxa are chosen to achieve roughly even coverage of the probable tree. Highly homoplastic changes may be extremely useful for resolving relationships among close relatives, and downweighting these would reduce confidence in these nodes.
Detecting Problems with the Quality of DNA Data
One surprising observation from our analysis is that an unexpectedly large number of slowly evolving characters only exhibited variation on terminal or subterminal branches. This was apparent from the large number of uninformative characters in slowly evolving partitions (Table 1) and also from the plot of average variability with respect to node level (Fig. 4). We can only guess what processes would produce a pattern contrary to the expectation of roughly proportionate levels of rate variation throughout the tree. One possibility is sequence error: the generation of a large DNA sequence database, from different gene regions and produced with different technologies (including cloning of PCR fragments in the earlier parts of this work; Vogler and Welsh, 1997) always bears a risk of sample mix-up and mistaken base calls. Because the majority of positions sequenced are invariant across our sample of taxa, a misread base pair would easily result in an autapomorphy for a single taxon; i.e., a base change at a terminal branch. For example, if misread bases were evenly distributed across the sequence and among taxa, and assuming our alignment includes around 1000 truly invariant characters, the observed number of 216 uninformative character changes could be obtained by a sequencing error rate of around 0.2% (= 216/(132 x 1000)). Simulations using SeqGen (Rambaut and Grassly, 1997) (by randomly exchanging bases in the data matrices simulated by SeqGen) confirmed that this could explain the excess of slow characters changing near the tips in Figure 4 (results not shown). Similar methods could be used to verify the quality of DNA data in systematic studies generally. Although this kind of error might have little effect on the topology, it could influence estimates of divergence times from sequence data which rely on autapomorphies as well as synapomorphies.
|
Conclusions |
|---|
|
TopAbstractMaterial and MethodsResultsDiscussionConclusionsAppendix 2AcknowledgmentsReferences |
|---|
With some 75% of all species-level taxa of Cicindela north of Mexico included here, a comprehensive assessment of relationships in this highly diverse genus is now possible, and necessitates several formal taxonomic changes (Duran and Vogler, 2005). Molecular systematics studies of this scale at the species level are still rare, in particular in insects, and the data can now be used to address various questions about speciation and diversification. We could distinguish four major clades corresponding to radiations confined to North America which are amenable for comparative studies of diversification rates within the continent and with other similar lineages of Cicindela (s.l.) elsewhere. However, before such studies can be performed, it will be necessary to overcome the difficulty of resolving the deeper nodes of the tree. Clearly this will not be achieved with downweighting the most highly variable characters. As our analysis has shown, the character distribution is such that homoplasy weighting reduced overall support. It would unnecessarily remove data with favorable properties, such as low rate heterogeneity along the root-to-tip axis and among different branches of the tree, and a high level of consistency with the tree locally. The effect of removing these data may also be that an incorrect topology is favored due to the diminished interaction of diverse types of characters. Whether our findings are specific to this particular data set, or generally to "species-level" trees with their great variance of divergences among taxa, is unclear. However, recent analyses using parsimony have generally supported the value of homoplastic variation for phylogenetic reconstruction (Kallersjo et al., 1999; Wenzel and Siddall, 1999). Similarly, model-based approaches may be very appropriate for analyzing questions about the information content of homoplastic variation in a statistical framework. Studies on simulated (Yang, 1998) or small empirical data sets (Goldman, 1998) generally agree with our conclusions from the parsimony analysis about the conditions under which highly homoplastic character variation can produce a meaningful phylogenetic signal. Finally, our results have demonstrated that even closely related species can show variation in the properties of nucleotides from one clade to another (covarion shifts, or heterotachy). Standard phylogenetic analysis methods assume that covarion shifts do not occur and thus can be misled. However, in the case of cicindelids, shifts in patterns of character variation appear to be weak and seem not to affect phylogenetic inferences, in particular due to the less biased 3rd-position characters.
|
|
Appendix 2 |
|---|
|
TopAbstractMaterial and MethodsResultsDiscussionConclusionsAppendix 2AcknowledgmentsReferences |
|---|
Topology of tree derived from LogDet distances. Note that major clades are similar to the parsimony tree, and differences to the equally weighted parsimony trees affect mostly taxa whose position had already been found to be uncertain under alternative weighting schemes (Table 4). For example, the three species of the aurora group in clade 2 are placed in the "red abdomen" group, and clade 4 includes the genus Microthylax which is morphologically and ecologically (salt flats and ocean beaches) similar to the remainder of this group. Interestingly, the divergent Ci. trifasciata is removed from clade 4 and is sister to clade 1 + clade 2, a position suggested also from the traditional taxonomy.
|
|
Acknowledgments |
|---|
|
TopAbstractMaterial and MethodsResultsDiscussionConclusionsAppendix 2AcknowledgmentsReferences |
|---|
We are grateful to D. Broszka, D. Pearson, and J. Stamatov for specimens. This work was funded by grant NERC/A/S/2000/00489 to APV and TGB and a fellowship of Fundação para a Ciência e a Tecnologia (PRAXIS XXI/BD/18409/98) to ACD. TGB is a Royal Society University Research Fellow. We thank Z. Yang, N. Takezaki, J. Bielawski, and C. Simon for many insightful comments and editorial help, and J. Pons, P. Foster, and D. Duran for discussions.
|
References |
|---|
|
TopAbstractMaterial and MethodsResultsDiscussionConclusionsAppendix 2AcknowledgmentsReferences |
|---|
-
Baker R. H., Wilkinson G. S., DeSalle R. Phylogenetic utility of different types of molecular data used to infer evolutionary relationships among stalk-eyed flies (Diopsidae). Syst. Biol. (2001) 50:87–105.
Barker F. K., Lutzoni F. M. The utility of the incongruence length differenc. test. Syst. Biol. (2002) 51:625–637.
Barraclough T. G., Hogan J. E., Vogler A. P. Testing whether ecological factors promote cladogenesis in a group of tiger beetles (Coleoptera: Cicindelidae). Proc. Roy. Soc. Ser. B (1999) 266:1061–1067.
Barraclough T. G., Nee S. Phylogenetics and speciation. Trend. Ecol. Evol. (2001) 16:391–399.[CrossRef][Medline]
Barraclough T. G., Vogler A. P. Recent diversification rates in North American tiger beetles (genus Cicindela). Mol. Biol. Evol. (2002) 19:1706–1716.
Börklund M. Are third positions really that bad? A test using vertebrate cytochrome b. Cladistics (1999) 15:191–195.[CrossRef][Web of Science]
Boyd H. P. Checklist of Cicindelidae: The tiger beetles (1982) Marlton, New Jersey: Plexus Publishing, Inc.
Cognato A. I., Vogler A. P. Exploring data interaction and nucleotide alignment in a multiple gene analysis of Ips (Coleoptera: Scolytinae). Syst. Biol. (2001) 50:758–780.
Collins T. M., Kraus F., Estabrook G. Compositional effects and weighting of nucleotide-sequences for phylogeneti. analysis. Syst. Biol. (1994) 43:449–459.
Cunningham C. W. Is congruence between data partitions a reliable predictor of phylogenetic accuracy? Empirically testing an iterative procedure for choosing among phylogeneti. methods. Syst. Biol. (1997) 46:464–478.
Darlu P., Lecointre G. When does the incongruence length difference test fail? Mol. Biol. Evol. (2002) 19:432–437.
DeBry R. W. Improving interpretation of the decay index for DNA sequenc. data. Syst. Biol. (2001) 50:742–752.
Dolphin K., Belshaw R., Orme C. D. L., Quicke D. J. L. Noise and incongruence: Interpreting results of the incongruence length differenc. test. Mol. Phylogenet. Evol. (2000) 17:401–406.[CrossRef]
Duran D. P., Vogler A. P. Taxonomic changes in North American Cicindela (Coleoptera: Cicindelidae). Coleopterists Bull. (2005) In preparation.
Farris J. S. A successive approximations approach to characte. weighting. Syst. Zool. (1969) 18:374–385.
Farris J. S., Källersjö M., Kluge A. G., Bult C. Testing significance of incongruence. Cladistics (1994) 10:315–320.[CrossRef][Web of Science]
Fitch W. F. The nonidentity of invariable positions in the cytochrome c of differen. species. Biochem. Genet. (1971) 5:231–241.[CrossRef][Web of Science][Medline]
Freitag R. A revision of the North American species of the Cicindela maritima group with a study of hybridization between Cicindela duodecimguttata and oregona. Quaest. Entomol. (1965) 1:87–170.
Gatesy J., O'Grady P., Baker R. H. Corroboration among data sets in simultaneous analysis: Hidden support for phylogenetic relationships among higher level Artiodactyl taxa. Cladistics (1999) 15:271–313.[CrossRef][Web of Science]
Goldman N. Phylogenetic information and experimental design in molecular systematics. Proc. Roy. Soc. Ser. B (1998) 265:1779–1786.
Goloboff P. A. Estimating character weights during tree search. Cladistics (1993) 9:83–91.[CrossRef][Web of Science]
Kallersjo M., Albert V. A., Farris J. S. Homoplasty increases phylogenetic structure. Cladistics (1999) 15:91–93.[CrossRef][Web of Science]
Lockhart P. J., Huson D., Maier U., Fraunholz M. J., Peer Y. V. D., Barbrook A. C., Howe C. J., Steel M. A. How molecules evolve in eubacteria. Mol. Biol. Evol. (2000) 17:835–838.
Mitchell A., Mitter C., Regier J. C. More taxa or more characters revisited: Combining data from nuclear protein-encoding genes for phylogenetic analyses of Noctuoidea (Insecta: Lepidoptera). Syst. Biol. (2000) 49:202–224.
Miyamoto M. M., Fitch W. M. Testing the covarion hypothesis of molecular evolution. Mol. Biol. Evol. (1995) 12:503–513.[Abstract]
Pearson D. L., Barraclough T. G., Vogler A. P. Distributional maps for North American species of tiger beetles. Cicindela (1997) 29:33–84.
Pearson D. L., Blum M. S., Jones T. H., Fales H. M., Gonda E., White B. R. Historical perspective and the interpretation of ecological patterns: Defensive compounds of tiger beetles (Coleoptera: Cicindelidae). Am. Nat. (1988) 132:404–416.[CrossRef][Web of Science]
Pearson D. L., Vogler A. P. Tiger beetles: The ecology and evolution of the Cicindelidae (2001) Ithaca, New York: Cornell University Press.
Philippe H., Casane D., Gribaldo S., Lopez P., Meunier J. Heterotachy and functional shift in protein evolution. IUBMB Life (2003) 55:257–265.[Web of Science][Medline]
Rambaut A., Grassly N. C. Seq-Gen: An application for the Monte Carlo simulation of DNA sequence evolution along phylogenetic trees. Comput. Applic. Biosci. (1997) 13:235–238.
Rivalier E. Démembrement du genre Cicindela Linné. Rev. Franc. Entomol. (1950) 17:217–244.
Rivalier E. Démembrement du genre Cicindela Linné. II. Faune americaine. Rev. Franc. Entomol. (1954) 21:249–268.
Rivalier E. Démembrement du genre Cicindela Linné. V. Faune australienne. Rev. Franc. Entomol. (1963) 30:30–48.
Sokal R. R., Rohlf F. J. Biometry. In: The principles and practice of statistics in biological research. (1995) 3rd edition. New York: W. H. Freeman.
Sorenson M. D. (1999) Boston, MA: Boston University. Treerot version 2.0.
Sota T., Vogler A. P. Incongruence of mitochondrial and nuclear gene trees in the carabid beetles Ohomopterus. Syst. Biol. (2001) 50:39–59.
Swofford D. L., Olson G. J., Waddell P. J., Hillis D. M. Phylogenetic inference. In: Molecular Systematics—Hillis D. M., Moritz C., Mable B. K., eds. (1996) Sunderland, Massachusetts: Sinauer. 407–514.
Vogler A. P., Goldstein P. Z. Adaptation, cladogenesis, and the evolution of habitat association in North American tiger beetles: A phylogenetic perspective. In: Molecular evolution and adaptive radiation—Givnish T., Systma K., eds. (1997) Cambridge, UK: Cambridge University Press. 353–373.
Vogler A. P., Kelley C. K. Covariation of defensive traits in Cicindela tiger beetles: A phylogenetic approach usin. mtDNA. Evolution (1998) 52:529–538.[CrossRef][Web of Science]
Vogler A. P., Welsh A. Phylogeny of North American Cicindela tiger beetles inferred from multiple mitochondrial DNA sequences. Mol. Phylogenet. Evol. (1997) 8:225–235.[CrossRef][Web of Science][Medline]
Vogler A. P., Welsh A., Barraclough T. G. Molecular phylogeny of the Cicindelamaritima group (Coleoptera: Cicindelidae) indicates fast radiation in western North America. Ann. Entomol. Soc. Am. (1998) 91:185–194.[Web of Science]
Walsh H. E., Kidd M. G., Moum T., Friesen V. L. Polytomies and the power of phylogenetic inference. Evolution (1999) 53:932–937.[CrossRef][Web of Science]
Wenzel J. W., Siddall M. E. Noise. Cladistics (1999) 15:51–64.[CrossRef][Web of Science]
Willis H. L. Bionomics and zoogeography of tiger beetles of saline habitats in the central United States (Coleoptera: Cicindelidae). Univ. Kans. Sci. Bull. (1967) 47:145–313.
Yang Z. H. On the best evolutionary rate for phylogeneti. analysis. Syst. Biol. (1998) 47:125–133.
Yoder A. D., Irwin J. A., Payseur B. A. Failure of the ILD to determine data combinability for slow loris phylogeny. Syst. Biol. (2001) 50:408–424.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  J. Pons, T. G. Barraclough, J. Gomez-Zurita, A. Cardoso, D. P. Duran, S. Hazell, S. Kamoun, W. D. Sumlin, and A. P. Vogler Sequence-Based Species Delimitation for the DNA Taxonomy of Undescribed Insects Syst Biol, August 1, 2006; 55(4): 595 - 609. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 W. Hordijk and O. Gascuel Improving the efficiency of SPR moves in phylogenetic tree search methods based on maximum likelihood Bioinformatics, December 15, 2005; 21(24): 4338 - 4347. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. Blaxter, J. Mann, T. Chapman, F. Thomas, C. Whitton, R. Floyd, and E. Abebe Defining operational taxonomic units using DNA barcode data Phil Trans R Soc B, October 29, 2005; 360(1462): 1935 - 1943. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||