© 2005 Society of Systematic Biologists
Untangling Long Branches: Identifying Conflicting Phylogenetic Signals Using Spectral Analysis, Neighbor-Net, and Consensus Networks
Edited by Olaf Bininda-Emonds: Associate Editor
1 Allan Wilson Centre for Molecular Ecology and Evolution, Department of Zoology, University of Otago P.O. Box 56, Dunedin, New Zealand E-mail: martyn.kennedy{at}stonebow.otago.ac.nz (M.K.)
2 Allan Wilson Centre for Molecular Ecology and Evolution, Institute of Fundamental Sciences, Massey University Private Bag 11222, Palmerston North, New Zealand
3 Department of Psychology, University of Auckland Private Bag 92019 Auckland, New Zealand
 |
Abstract |
|---|
 Top Abstract MethodsResultsDiscussionAcknowledgmentsReferences |
|---|
Long-branch attraction is a well-known source of systematic error that can mislead phylogenetic methods; it is frequently invoked post hoc, upon recovering a different tree from the one expected based on prior evidence. We demonstrate that methods that do not force the data onto a single tree, such as spectral analysis, Neighbor-Net, and consensus networks, can be used to detect conflicting signals within the data, including those caused by long-branch attraction. We illustrate this approach using a set of taxa from three unambiguously monophyletic families within the Pelecaniformes: the darters, the cormorants and shags, and the gannets and boobies. These three families are universally acknowledged as forming a monophyletic group, but the relationship between the families remains contentious. Using sequence data from three mitochondrial genes (12S, ATPase 6, and ATPase 8) we demonstrate that the relationship between these three families is difficult to resolve because they are separated by a short internal branch and there are conflicting signals due to long-branch attraction, which are confounded with nonhomogeneous sequence evolution across the different genes. Spectral analysis, Neighbor-Net, and consensus networks reveal conflicting signals regarding the placement of one of the darters, with support found for darter monophyly, but also support for a conflicting grouping with the outgroup, pelicans. Furthermore, parsimony and maximum-likelihood analyses produced different trees, with one of the two most parsimonious trees not supporting the monophyly of the darters. Monte Carlo simulations, however, were not sensitive enough to reveal long-branch attraction unless the branches are longer than those actually observed. These results indicate that spectral analysis, Neighbor-Net, and consensus networks offer a powerful approach to detecting and understanding the source of conflicting signals within phylogenetic data.
Keywords: Anhingidae; consensus networks; darters; long-branch attraction; Neighbor-Net; Pelecaniformes; spectral analysis
Received August 30, 2004; Revised November 8, 2004; Accepted February 22, 2005
A phylogenetic approach has proved to be invaluable in many areas of comparative biology including the study of adaptation (e.g., Larson and Losos, 1996), behavior (e.g., Kennedy et al., 1996; Slikas, 1998), biogeography (e.g., Nelson and Platnick, 1981; Cracraft, 1994), and coevolution (e.g., Hafner et al., 1994; Paterson and Gray, 1996; Page et al., 1998). The usefulness of such studies is limited, however, by the accuracy of the phylogenies on which they are based. Finding reliable estimates of species phylogeny is a long-studied and challenging problem (Felsenstein, 2003; Semple and Steel, 2003). Aside from sampling error, there are several sources of systematic error that can mislead phylogenetic methods, such as varying substitution processes across the tree. Also, some types of tree are harder to infer correctly. In particular, the presence of multiple long branches in the tree is known to mislead some methods (Felsenstein, 1978; Hendy and Penny, 1989). Here we explore this particular type of systematic error, known as "long-branch attraction," in the context of the phylogeny of the Pelecaniformes. Our main interest is in how this and other sorts of conflicting phylogenetic signal can be detected using a range of exploratory data analysis tools.
Although long-branch attraction (LBA) may be suspected if two long branches are placed together on a phylogenetic tree, it should not be automatically invoked as the branches may truly belong together (Huelsenbeck, 1998; Wiens and Hollingsworth, 2000). Thus, LBA cannot be used to explain a problematic phylogeny without further justification. Current approaches used to diagnose a LBA problem are applied after the phylogenetic analysis has been performed and are not usually applied unless the estimated tree is in conflict with a prior hypothesis. LBA is often used as an explanation if the outgroup (frequently a long branch) roots the tree in an unexpected way or if other long-branch taxa within the ingroup appear to be drawn towards the root away from their expected position (for instance, it has been suspected that LBA draws the hedgehog towards the root of the placental mammals, Penny et al., 1999; Lin et al., 2002). Further evidence of a LBA is found if the parsimony tree is different from the maximum-likelihood tree, as the parsimony method has been shown to be particularly vulnerable to LBA effects (Felsenstein, 1978; Hendy and Penny, 1989). One statistical test that has been proposed to test for LBA is due to Huelsenbeck (1997; 1998), who described a method using Monte Carlo simulation (the parametric bootstrap) that generates multiple data sets (from a parameter-rich model of DNA substitution) on a tree in which the long branches are separated. These simulated data sets are then analyzed using parsimony to find what proportion of the time the long branches are placed together.
The Monte Carlo simulation method can only be applied if LBA is suspected on the basis of the phylogenetic analysis. It would be ideal if a preliminary analysis of the data could show the presence of conflicting signals, and alert the systematist to the potential for LBA. Fortunately, phylogenetic tools capable of doing just this are available. In this paper we show how spectral analysis (Hendy and Penny, 1993), NeighborNet (Bryant and Moulton, 2004), and consensus networks (Holland and Moulton, 2003; Holland et al., 2005) can be used to detect conflicting signals in the data, and hence help infer whether or not long branches are attracted to one another. These three methods are all split-based (a split is a bipartition of the taxa set that corresponds to a branch in a tree), where the weight of a split indicates how much support there is for that bipartition. (Note that we use the term support in the sense used in Wilkinson et al. [2003] and see no contradiction in saying that conflicting splits can be supported by the data.) There are other methods designed to assess the "treelikeness" of data that are not based on splits (e.g., delta-plots [Holland et al., 2002]), as well as techniques that attempt to detect LBA (e.g., RASA plots [Lyons-Weiler and Hoelzer, 1997]).
Spectral analysis evaluates the phylogenetic information in sequence data independently of a tree topology (Hendy and Penny, 1993; Lockhart et al., 1995) by computing the level of support and conflict for all possible branches of the tree (for an introduction to spectral analysis see Charleston and Page, 1999). If long-branch attraction is a problem, spectral analysis should indicate that there is support for two or more conflicting (i.e., mutually exclusive) splits, one of which groups the long branches together. Thus, because the splits are mutually exclusive, they must simultaneously have high conflict (presuming they each have high support). Neighbor-Net (Bryant and Moulton, 2004) is an agglomerative method related to neighbor-joining (Saitou and Nei, 1987) and split decomposition (Bandelt and Dress, 1992) that produces phylogenetic networks. Neighbor-Nets are often more resolved than split-decomposition networks, especially when the number of taxa is large. In a Neighbor-Net, contradictory signals are represented by box-like parts of the graph, whereas portions of the graph with little conflict appear more tree-like. Consensus networks are a generalization of consensus trees that display all branches that appear in an input set of trees above some frequency threshold. For example, one could display all branches that appeared in more than 10% of the trees generated by a bootstrap analysis. The most important difference between consensus networks and the other methods is that they begin with a set of trees as the input, rather than a sequence alignment or distance matrix. Neighbor-Nets and consensus networks are both capable of displaying conflicting evolutionary hypotheses. Long-branch attraction should, therefore, produce box-like regions in the graphs where the long branches are present. The box-like regions will reflect the support for the "true" relationship with one set of parallel edges and the support for any conflicting long-branch attraction signal with another set of parallel edges.
To evaluate whether spectral analysis, Neighbor-Net, and consensus networks can be used to infer LBA we use a data set containing members of the Pelecaniformes (Table 1). The pelecaniforms have a varied taxonomic history and offer some of the more interesting questions in avian phylogeny (Sibley and Ahlquist, 1990). Within the pelecaniforms, three clearly defined and unambiguously monophyletic families (the Anhingidae, Phalacrocoracidae, and Sulidae) form a monophyletic group (see Cracraft, 1985; Sibley and Ahlquist, 1990; Hedges and Sibley, 1994; Kennedy et al., 1996; Siegel-Causey, 1997; Kennedy and Spencer, 2004), but the relationship between these families remains uncertain. Cracraft (1985) and Siegel-Causey (1997) place the Anhingidae (darters) as sister taxa to the Phalacrocoracidae (cormorants and shags), whereas Sibley and Ahlquist's (1990) UPGMA tree and taxonomy place the Anhingidae as sister taxa to the Sulidae (boobies and gannets) (also see van Tuinen et al., 2001; Kennedy and Spencer, 2004).
|
|
Methods |
|---|
|
TopAbstractMethodsResultsDiscussionAcknowledgmentsReferences |
|---|
Sequence Data
Total genomic DNA was obtained for each of the samples using proteinase K followed by phenol/chloroform extraction. Once extracted the DNA was amplified for three mtDNA genes, the 12S ribosomal RNA gene and the ATPase 6 and ATPase 8 genes. The polymerase chain reaction (PCR) was used to amplify these regions using universal primers for 12S (Kocher et al., 1989) and primers for ATPase 6 and ATPase 8 (see Kennedy and Spencer, 2004). PCR amplification followed protocols described previously (see Kennedy et al., 2000; Kennedy and Spencer, 2000, 2004). The PCR product was purified using Gelase (Epicentre Tech.) and then sequenced by an automated sequencer using either the PCR primers or internal primers (see Kennedy and Spencer, 2004). The short ATPase 8 fragment (137 bp) could not be sequenced for A. novaehollandiae. Wherever possible both strands of DNA were sequenced for two or more individuals of a species to verify the accuracy of the sequencing and control for DNA contamination. The species for which only one individual was available were amplified and sequenced more than once to ensure consistent results. On those occasions where it was not possible to discriminate between alternative bases at a site IUPAC ambiguity codes were used.
Phylogenetic Analyses
Alignment procedures have been described in Kennedy et al. (2000). The alignment gave a 383-bp fragment of 12S, whereas the overlapping ATPase 6 and ATPase 8 coding genes gave a 758-bp fragment. The sequences have been submitted to GenBank (Accession numbers: AY009321
[GenBank]
–AY009323
[GenBank]
, AY009333
[GenBank]
, AY009338
[GenBank]
, AY009345
[GenBank]
–AY009347
[GenBank]
, AY009357
[GenBank]
, AY009362
[GenBank]
, AY369047
[GenBank]
, AY369048
[GenBank]
, AY369071
[GenBank]
, AY369072
[GenBank]
, AY941805
[GenBank]
–AY941812
[GenBank]
) and the aligned data matrix and resultant phylogenetic trees have been submitted to TreeBASE (www.treebase.org; S1257 & M2196).
We used the partition-homogeneity test (Swofford, 2002; the ILD of Farris et al., 1995) to investigate whether the ATPase and 12S sequences can be analyzed as a single data set, and the PTP test (Faith, 1991; Faith and Cranston, 1991) and g1 statistic (Hillis and Huelsenbeck, 1992) to investigate whether the data contained significant signal.
The program Spectrum 2.3 (Charleston, 1998) was used to perform spectral analysis (Hendy and Penny, 1993). In spectral analysis support for a split depends on the number of character columns in the alignment whose patterns correspond to that split, whereas the conflict for a split is the sum of the support for the splits that are incompatible with it. As a split may be incompatible with many other splits, its conflict may be much larger than its support. To make the level of conflict comparable to the level of support the conflict values are normalized (see Lento et al., 1995). Spectrum computes the support and conflict for all the terminal and possible internal branches; a threshold (in this case 0.001) is used to avoid calculating extremely low, biologically irrelevant, support values. The resulting spectrum is plotted as a bar chart (see Lento et al., 1995), which allows the level of support and conflict for the internal (possibly mutually exclusive) branches of interest to be visually compared.
The program SplitsTree4 (also known as jSplits, available at www-ab.informatik.uni-tuebingen.de/software/jsplits/welcome_en.html) was used to generate the Neighbor-Nets. For this analysis we used uncorrected distances (also known as p-distances or Hamming distances). We used this option for both the Neighbor-Nets and spectral analysis as it is the most similar to equally weighted parsimony, and therefore the most likely to reveal any conflicting signals that could mislead a parsimony analysis.
Parsimony and maximum-likelihood analyses were conducted using version 4.0b10 of PAUP* (Swofford, 2002). For the parsimony optimality criterion we used the branch-and-bound option in PAUP*, which guarantees to find the optimal tree. For the maximum likelihood optimality criterion we used heuristic search with TBR branch-swapping, the starting tree was obtained by stepwise addition using the as-is addition order. The model of nucleotide substitution used by maximum likelihood was selected using Modeltest (Posada and Crandall, 1998) under the Akaike Information Criterion. The model chosen was the general time-reversible model with a gamma distribution of rates (four categories) and a proportion of invariant sites. The parameters for this model, including base frequencies, were all estimated by maximum likelihood.
Two sets of 1000 bootstrap replicates (Felsenstein, 1985) were created for each of parsimony and maximum likelihood, using a full-heuristic search in each case. For the maximum-likelihood bootstrap the parameters of the GTR+I+G model were fixed at the values estimated by maximum likelihood. The bootstrap analyses were visualized using consensus networks (Holland and Moulton, 2003; Holland et al., 2005) with a threshold of 10%. That is, all edges that appeared in more than 10% of the bootstrap replicates were displayed in the network. Spectronet (Huber et al., 2002) was used to draw the consensus networks.
MrBayes v3.0 (Ronquist and Huelsenbeck, 2003) was used for Markov chain Monte Carlo Bayesian posterior probabilities with the following settings: the maximum-likelihood model employed 6 substitution types ("nst = 6") and rate variation across sites was modeled using a gamma distribution, with a proportion of the sites being invariant ("rates = invgamma"). The Markov chain Monte Carlo search was run with 4 chains for 2,000,000 generations, with trees being sampled every 100 generations (the first 5,000 trees, i.e., 500,000 generations, were discarded as "burnin").
The test for long branch attraction described in Huelsenbeck (1998) was implemented using Seq-Gen v1.2.5 (Rambaut and Grassly, 1997) to simulate multiple data sets. Parsimony trees for these data sets were then constructed using PAUP* as described above.
As studies by Darlu and Lecoinre (2002) and Dolphin et al. (2000) have questioned the power of the ILD test of Farris et al. (1995) to detect heterogeneity, we also repeated the phylogenetic analyses described above for the ATPase and 12S data sets individually.
|
Results |
|---|
|
TopAbstractMethodsResultsDiscussionAcknowledgmentsReferences |
|---|
The partition-homogeneity test found no significant difference in phylogenetic signal between the ATPase and 12S sequences and thus suggest that the sequences could be analyzed as a single data set (100 replicates, P = 0.28). Of the 450 variable sites, some 313 of the characters were parsimony informative. Both the significantly skewed tree-length distribution (g1 = –1.155 from 10000 random trees, P < 0.01; Hillis and Huelsenbeck, 1992) and a PTP test (1000 replicates, P = 0.001) showed that the data contain significant signal.
To investigate evidence for conflicting signals in our data independent of any tree we used spectral analysis and Neighbor-Nets. Spectral analysis (Fig. 1) supports the monophyly of each of the major groups: Phalacrocoracidae, Sulidae, and Anhingidae. The split labels given in Figure 1 identify the potential clades of interest; this labeling of splits is used in the trees and networks that follow. Support for the Phalacrocoracidae as a group (split B) is high (0.0384; the units are expected number of substitutions per site), and the normalized conflict is low (0.0029). Similarly, support for the Sulidae as a group (split G) is high (0.0105) in comparison with its conflict (0.0032). The monophyly of the darters (split F) is well supported (0.0114), but also has relatively high conflict (0.0073). The source of this conflict includes the support for an alternative split (K, support = 0.0047), which groups the Pelecanidae and A. anhinga and the support for the split (L, support = 0.0033) that groups the darters and Pe. conspicillatus. Of the higher level groupings, support for Anhingidae + Pelecanidae (split H) is relatively low (0.0079) compared with its conflict (0.0061). Support for the alternative Anhingidae + Sulidae (split N) is low (0.0005) compared with its conflict (0.0143).
|
Neighbor-Nets (Bryant and Moulton, 2004) were used to visualize conflicting signals in the data and to assess the impact of the inclusion of A. anhinga (Fig. 2). The network of the complete data set (Fig. 2a) shows considerable evidence of conflict concerning the placement of A. anhinga and the monophyly of the darters. Both split F (darter monophyly) and split K (the grouping of A. anhinga with the pelicans) are found in the network with roughly equal support. There is also support for the grouping of A. anhinga and A. rufa (split I). Although there is strong support for the monophyly of the pelicans (split A), there is also some support for a group of Pe. conspicillatus with the darters (split L). If we exclude A. anhinga from the Neighbor-Net analysis the conflicting signal within the darters is significantly reduced (Fig. 2b). If A. anhinga is the only darter included in the Neighbor-Net (Fig. 2c), it groups with the pelicans. The only other substantial amount of conflict shown by the Neighbor-Nets concerns the position of the Australasian Gannet (M. serrator). M. serrator joins the graph near the central polytomy, but the network shows that there is some contradictory signal linking it to S. sula(Fig. 2a, split J). The network (Fig. 2a) shows significant support for the monophyly of Pelecanidae (split A) and Phalacrocoracidae (split B), slight support for the monophyly Anhingidae (split F), and little support for the monophyly of Sulidae. There is no support, however, for any of the three alternative groupings of the families Anhingidae, Phalacrocoracidae, and Sulidae.
|
Our equally weighted parsimony analysis found two most parsimonious trees, which differ only in the monophyly of the darters (Fig. 3). The monophyly of the darters (split F) is uncontroversial (see Johnsgard, 1993), but it received low bootstrap support (55%, Fig. 3b). Given that the darters are known to be monophyletic, their lack of monophyly in Figure 3a may represent LBA between the long terminal-branch of A. anhinga and long internal-branch that groups the outgroup pelicans. The accepted monophyly of the darters fulfils one of Wiens and Hollingsworth's (2000) suggested criteria for inferring LBA: that the long branches are known not to be sister taxa (i.e., A. anhinga is known to be a darter and not a pelican from extensive behavioral and morphological evidence, see Johnsgard, 1993).
|
The consensus network of the 1000 parsimony bootstrap replicates is shown in Figure 4. It displays two major conflicts, the biggest box in the network reflects the uncertainty over the placement of A. anhinga, the two options, of either monophyletic darters (split F) or A. anhinga grouping with the pelicans (split K), have similar bootstrap values (55% and 45%, respectively). The second conflict is whether Phalacrocoracidae + Sulidae (split H, bootstrap value of 72%) form a sister group, or if Anhingidae + Sulidae (split N, bootstrap value of 20%) form a sister group.
|
Using the GTR+I+G model selected by Modeltest (Posada and Crandall, 1998), we ran a maximum-likelihood analysis estimating all parameters of the model. The estimated model parameters were: base frequencies (0.3249, 0.3661, 0.1227, 0.1873); substitution model [A-C] 3.8257, [A-G] 30.3249, [A-T] 6.2428, [C-G] 1.7799, [C-T] 53.2645, [G-T] 1.0000; proportion of invariable sites 0.4720; gamma distribution shape parameter 1.0614. The maximum-likelihood tree (Fig. 5) places the Anhingidae as sister taxa to the Sulidae, in contrast to either of the parsimony trees, although it is worth noting that the consensus network of the parsimony bootstrap replicates also indicates some signal for the Anhingidae + Sulidae group (split N). The Bayesian analysis also recovered the maximum-likelihood topology, as did maximum parsimony with transversions weighted
3 times transitions (results not shown).
|
The consensus network of the 1000 ML bootstrap replicates (Fig. 6) shows that of the three possible sister taxa arrangements for the Anhingidae, Phalacrocoracidae, and Sulidae, the best supported is Anhingidae + Sulidae (split N, bootstrap value of 51%), followed by Phalacrocoracidae + Sulidae (split H, bootstrap value of 38%), the third possibility, Anhingidae + Phalacrocoracidae, receives bootstrap support of 11%. This result suggests that longer sequences will be required to unequivocally resolve the relationship for this short internal branch. In contrast to the parsimony analysis, there is no signal placing A. anhinga with the pelicans in the consensus network of the maximum-likelihood bootstrap replicates.
|
Evidence from the spectral analysis, Neighbor-Nets, maximum-likelihood and parsimony trees, and consensus networks all suggests that LBA between the darters (especially A. anhinga) and the pelicans (splits K and L) may have pulled the Anhingidae as a whole towards the pelicans in the parsimony trees. To test this hypothesis, the data were reanalyzed with A. anhinga excluded to negate the effect of it being attracted to either the other darters (split F) or the pelicans (split K). Figure 7 shows the consensus networks for 1000 bootstrap replicates using both parsimony (Fig. 7a) and maximum likelihood (Fig. 7b); branches in the parsimony and maximum-likelihood trees are shown in bold. The absence of A. anhinga makes an obvious difference to the parsimony analysis and the level of bootstrap support for the placement of the darters. The parsimony tree and the maximum-likelihood tree with A. anhinga excluded are both congruent with the maximum-likelihood tree with it included (Fig. 5). Parsimony analysis with A. anhinga excluded places the remaining darters as sister taxa to the Sulidae in 60% of the bootstrap replicates (Fig. 7a). The maximum-likelihood bootstrap with A. anhinga excluded gave 81% support for the darters as sister taxa to the Sulidae (Fig. 7b). It appears that the inclusion of A. anhinga in this data set produces misleading results for parsimony because both A. anhinga and the other darters are drawn away from the Sulidae towards the pelicans (as seen in Figs. 3a, b). It is also interesting to note that when A. anhinga is excluded from the maximum-likelihood analysis, the bootstrap support for the Anhingidae + Sulidae grouping increases from 51% to 81%. This finding that the presence of A. anhinga greatly affects the inferences made is mirrored by the Bayesian posterior probabilities where support for grouping Anhingidae + Sulidae increases from 0.62 (when A. anhinga is included) to 0.96 when A. anhinga is excluded from the analysis (Fig. 8). All these results support our contention that the conflict in our data set is manifested in the grouping of the darters with the pelicans instead of the gannets and boobies.
|
|
The initial spectral and Neighbor-Net analyses showed evidence for conflicting signal within the data set. This conflict was further corroborated, first, by the differences between the maximum-likelihood and parsimony analyses; and second, by the observation that excluding A. anhinga resolves these differences. To see if there was also statistical evidence for LBA misleading the parsimony analysis we implemented the Monte Carlo simulation approach described in Huelsenbeck (1998).
The Monte Carlo simulations (Fig. 9) used the maximum-likelihood tree shown in Figure 5 and the ML estimates of the GTR+I+G parameters. The lengths of the long branches leading to the pelicans and A. anhinga were varied (0.1 to 0.5 in steps of 0.05) but were always equal to each other; all other branch lengths were left at their ML estimates. We generated 100 data sets each for a range of long-branch lengths. In a separate simulation, using the actual observed branch lengths, parsimony did not recover the maximum-likelihood tree 23 times out of 100, but only 2 of these 23 incorrect trees contained the group of A. anhinga with the pelicans (split K, as found in the parsimony tree, Fig. 3a). The proportion of simulated data sets in which the maximum-likelihood topology is not recovered is quite high even when the long-branch lengths are less than 0.3 substitutions per site; however, monophyly of the darters was still observed in more than 95% of simulations until branches were longer than 0.3 substitutions per site. Although the actual long-branch lengths (0.198 and 0.223) are within the range in which the simulations produce inaccurate trees (e.g., they may cause the darters as a whole to be attracted to the pelicans), these inaccurate trees infrequently contained the split K found in the parsimony analysis (Fig. 3a).
|
Despite the results of the ILD test indicating that the genes could be treated as a single data set, we performed some analysis of the ATPase and 12S genes separately. The ML and MP trees estimated for ATPase were the same as the tree shown in Figure 5, but the ML and MP trees estimated for 12S were the same as the tree shown in Figure 3a (i.e., they display the suspected LBA). Spectral analysis (Fig. 10), consensus network analysis of MP bootstrap (results not shown), and Neighbor-Net analysis (results not shown) indicated that different splits received different support in the ATPase genes to the 12S gene. In particular split K linking A. anhinga to the pelicans was present in the 12S data but was absent in the ATPase data. However, both 12S and the ATPase data did have signal linking the darters as a whole to the pelicans (split H). Both genes showed conflicting signal that could be attributable to LBA but it was manifested in different ways.
|
|
Discussion |
|---|
|
TopAbstractMethodsResultsDiscussionAcknowledgmentsReferences |
|---|
Although long-branch attraction is a problem for phylogenetic reconstruction (Hendy and Penny, 1989; Swofford et al., 1996; Anderson and Swofford, 2004), it should not be used as an explanation of a problematic phylogeny whenever two long branches are placed together, as those branches may truly belong together. In this paper we advocate using methods that are capable of displaying conflicting phylogenetic signals, and thus capable of identifying the potential for LBA. By using methods such as spectral analysis, Neighbor-Net, and consensus networks we were able to show that there is signal for placing A. anhinga in more than one part of the Pelecaniformes tree. The existence of a long-branch attraction problem was corroborated by the differences between the parsimony and maximum-likelihood analyses. Maximum likelihood is known to be less sensitive to LBA than parsimony is (Huelsenbeck, 1997, 1998; Cunningham et al., 1998; Wiens and Hollingsworth, 2000; Holland et al., 2003), so we hypothesize that the difference between the likelihood and parsimony trees in this analysis is due to LBA between the terminal branch of A. anhinga and the internal branch of the pelicans.
Using a combination of methods to detect LBA, as described here, is a useful approach as the various methods have different strengths and weaknesses. Spectral analysis's main advantage is the ability to show both the support and the conflict for different groupings. The main disadvantage of spectral analysis is that the barchart display alone does not readily facilitate phylogenetic interpretation (although this problem can be overcome by labeling the splits on any trees or other graphical representations of the information). Neighbor-Net gave a useful graphical display of the conflicting signals in these data; it is limited to displaying 2-dimensional graphs, but for those data that seemed to capture almost all of the relevant signal. Unlike split-decomposition graphs (which can also be used to visualize conflicting signal [Clements et al., 2003]), Neighbor-Nets are not known to have a tendency to become star-like as more taxa are added. Consensus networks were useful for visualizing the results of the bootstrap analyses; their limitation for detecting LBA is that bootstrap values are designed to detect sampling error rather than systematic error. When systematic error is a problem (e.g., in the Felsenstein zone for parsimony analysis; Felsenstein, 1978), there could be 100% bootstrap support for an incorrect branch of the tree and the consensus network would not display any conflict. This drawback could be overcome to some extent by including the conflict values from spectral analysis in an annotation of the trees or consensus networks.
In addition to the methods discussed above, we used the nonparametric bootstrap test of Huelsenbeck (1998), which showed that parsimony could be misled for data generated on the maximum-likelihood tree (23 times out of 100 for data generated on a tree with the observed branch lengths). For the length of the long branches in this data set, Monte Carlo simulations demonstrate that the maximum-likelihood tree topology will not always be recovered, but they do not suggest that the branches are long enough to cause the non-monophyly of the darters as found in the parsimony analysis (Fig. 3a). The simulation approach appears to exhibit less sensitivity to the conflicting signals in the data than the alternative approaches: spectral analysis, Neighbor-Net, and consensus networks.
It would be unwise to make a strong recommendation about which methods are best to use on the basis of a single case study. Future work that investigates which methods are most effective at detecting LBA could include a simulation study on trees in which long branch effects have been intentionally introduced, and a wider range of case studies. In the meantime, as the methods have different strengths and weaknesses, and none are unduly computational expensive, we recommend using them all.
Two further notes of caution should be applied before interpreting splits graphs as showing LBA. First, LBA is not the only effect that will result in data containing conflicting signals. Complex nucleotide substitution processes where base composition, or the ability of sites to accept mutations, changes across the tree could also cause strong nonhistorical signals. In some data there may also be historical processes that lead to conflicting signal, for example, lineage sorting, hybridization, or recombination. Second, there may be cases where signal in the data caused by complex substitution processes completely overwhelms the true historical signal, so there would be no observable conflict.
The effects of long-branch attraction are not necessarily obvious in this data set, with the conflicting signal regarding the placement of A. anhinga causing the darters as a whole to be attracted to the pelicans. If our analyses had found a single parsimony tree with monophyletic darters (e.g., Fig. 3b), we may not have been alerted to the possible LBA problem in this data set. Our finding that long-branch attraction may manifest itself by moving a group of taxa across a short internal branch rather than simply by placing the long branches together may appear somewhat disconcerting (as it is a less obvious problem).
Several techniques have been recommended as ways to avoid long-branch attraction; here we showed that removing a long branch can be useful. We used splits-based exploratory data analysis tools to determine which taxon was responsible for much of the conflict but other approaches (often involving jackknifing) are available. For instance, Thorley and Wilkinson (1999) developed a measure of leaf instability that could potentially be used to detect long-branch taxa. The use of the reduced consensus methods (Wilkinson, 1996) might also detect that A. anhinga was responsible for a large proportion of the conflict. Alternatively, increased taxonomic sampling could be useful, as phylogeny estimation is generally thought to be aided by breaking up long branches (Swofford et al., 1996; Graybeal, 1998; Hillis, 1998; Poe, 1998; Anderson and Swofford, 2004). For example, increased taxon sampling of whole mitochondrial genomes recovers the traditional root of the avian tree rather than the suggested basal passerines (see Mindell et al., 1999; Harrison et al., 2004). Thus, if sequences from other pelicans or darters were added it might resolve the problem with this data set. Similarly, additional sequence data might also resolve the problem (although we note that this is offered as a panacea for most phylogenetic inference problems, and would not help in cases of pure LBA, in which increased convergence to the wrong tree would result). The effects of minor cases of long-branch attraction are probably readily alleviated by using corrections for multiple substitutions at a site. For example, in this data set maximum likelihood (and even simply increasing the relative weight of transversions under parsimony) will alleviate the affect of LBA to some extent.
Removing the conflict in our data set due to LBA, by excluding the taxon A. anhinga, increased the support for the Anhingidae + Sulidae internal branch under both parsimony and maximum-likelihood analyses. Excluding this taxon ameliorates the LBA problem and, while there is still contradictory signal linking Pe. conspicillatus to the darters, it is not sufficient to mislead the parsimony method. Nevertheless, the difficulty of lack of signal still remains; the small amount of change accrued during the diversification of the groups Anhingidae, Phalacrocoracidae, and Sulidae makes it difficult to unambiguously resolve the phylogenetic relationships of these families in this study regardless of the LBA issues. (Note, however, that there was little or no support for the traditional topology grouping the Anhingidae with the Phalacrocoracidae.) The consistency of parsimony is affected by the relative length of the short internal branches (like this one) compared with the other branches (Schulmeister, 2004), and thus additional data may change the relative length of the short internal branch, and hence alter the risk of inconsistency.
In summary, using the phylogeny of the Pelecaniformes as our example, we have demonstrated that exploratory data analysis tools, such as spectral analysis, Neighbor-Net, and consensus networks, make it possible to visualize conflicting signals within the Pelecaniformes data set. These signals appear to have two sources, long-branch attraction, and nonhomogenous sequence evolution across the ATPase and 12S genes. Long-branch attraction is a ubiquitous and sometimes cryptic source of systematic bias in phylogenetic estimation, but Monte Carlo methods to test for it can be applied only after LBA has been identified as a potential problem. To counteract this drawback we suggest using methods capable of displaying conflicting signal as data exploration tools alongside the standard tree estimation methods.
|
|
Acknowledgments |
|---|
|
TopAbstractMethodsResultsDiscussionAcknowledgmentsReferences |
|---|
We thank the following people and institutions for supplying samples: A. Baker and O. Haddrath (Royal Ontario Museum), L. Christidis (Museum of Victoria), D. Dittmann (Museum of Natural Science at Louisiana State University), B. Gill (Auckland Institute and Museum), C. Robertson (New Zealand Department of Conservation), and C. Wood (Burke Museum). O. Bininda-Emonds, M. Charleston, J. Cotton, J. Cracraft, D. Faith, F.-J. Lapointe, S. Greenhill, R. Olmstead, R. Page, D. Penny, M. Roy, C. Semple, and anonymous referees provided comments during preparation of the manuscript. The genetic work was carried out in G. Wallis' Evolutionary Genetics Laboratory and was assisted by the advice of T. King and L. Wallis. This work was supported by grants from the University of Otago, the University of Auckland, and the New Zealand Lotteries Board, Science Research Committee (to RDG and HGS). MK was supported by a University of Otago Postgraduate Scholarship and a New Zealand Science and Technology Postdoctoral Fellowship. BRH was supported by a New Zealand Science and Technology Postdoctoral Fellowship.
|
References |
|---|
|
TopAbstractMethodsResultsDiscussionAcknowledgmentsReferences |
|---|
-
Anderson F. E., Swofford D. L. Should we be worried about long-branch attraction in real data sets? Investigations using metazoan 18S rDNA. Mol. Phylogenet. Evol. (2004) 33:440–451.[CrossRef][Web of Science][Medline]
Bandelt H.-J., Dress A. W. M. A canonical decomposition theory for metrics on a finite set. Adv. Math. (1992) 92:47–105.[CrossRef]
Bryant D., Moulton V. Neighbor-Net: An agglomerative method for the construction of phylogenetic networks. Mol. Biol. Evol. (2004) 21:255–265.
Charleston M. A. Spectrum: Spectral analysis of phylogenetic data. Bioinformatics (1998) 14:98–99.
Charleston M. A., Page R. D. M. Spectral analysis—A brief introduction. In: Molecular systematics and plant evolution—Hollingsworth P. M., Bateman R. M., Gornall R. J., eds. (1999) London: Taylor & Francis. Pages 255–267.
Clements K. D., Gray R. D., Choat J. H. Rapid evolutionary divergences in reef fishes of the family Acanthuridae (Perciformes: Teleostei). Mol. Phylogenet. Evol. (2003) 26:190–201.[CrossRef][Web of Science][Medline]
Cracraft J. Monophyly and the phylogenetic relationships of the Pelecaniformes: A numerical cladistic analysis. Auk (1985) 102:834–853.[Web of Science]
Cracraft J. Species diversity, biogeography, and the evolution of biotas. Ame. Zool. (1994) 34:33–47.
Cunningham C. W., Zhu H., Hillis D. M. Best-fit maximum-likelihood models for phylogenetic inference: Empirical tests with known phylogenies. Evolution (1998) 52:978–987.[CrossRef][Web of Science]
Darlu P., Lecointre G. When does the incongruence length difference test fail? Mol. Biol. Evol. (2002) 19:432–437.
Dolphin K., Belshaw R., Orme C. D. L., Quicke D. L. J. Noise and incongruence: Interpreting results of the incongruence length difference test. Mol. Phylogenet. Evol. (2000) 17:401–406.[CrossRef][Web of Science][Medline]
Faith D. P. Cladistic tests for monophyly and nonmonophyly. Syst. Zool. (1991) 40:366–375.
Faith D. P., Cranston P. S. Could a cladogram this short have arisen by chance alone? On permutation tests for cladistic structure. Cladistics (1991) 7:1–28.[CrossRef][Web of Science]
Farris J. S., Källersjö M., Kluge A. G., Bult C. Constructing a significance test for incongruence. Syst. Biol. (1995) 44:570–572.
Felsenstein J. Cases in which parsimony or compatibility methods will be positively misleading. Syst. Zool. (1978) 27:401–410.
Felsenstein J. Confidence limits on phylogenies: An approach using the bootstrap. Evolution (1985) 39:783–791.[CrossRef][Web of Science]
Felsenstein J. Inferring phylogenies (2003) Sunderland, Massachusetts: Sinauer Associates.
Graybeal A. Is it better to add taxa or characters to a difficult phylogenetic problem? Syst. Biol. (1998) 47:9–17.
Hafner M. S., Sudman P. D., Villablanca F. X., Spradling T. A., Demastes J. W., Nadler S. A. Disparate rates of molecular evolution in cospeciating hosts and parasites. Science (1994) 265:1087–1090.
Harrison G. L., McLenachan P. A., Phillips M. J., Slack K. E., Cooper A., Penny D. Four new avian mitochondrial genomes help get to basic evolutionary questions in the Late Cretaceous. Mol. Biol. Evol. (2004) 21:974–983.
Hedges S. B., Sibley C. G. Molecules vs. morphology in avian evolution: The case of the "pelecaniform" birds. Proc. Natl. Acad. Sci. USA (1994) 91:9861–9865.
Hendy M. D., Penny D. A framework for the quantitative study of evolutionary trees. Syst. Zool. (1989) 38:297–309.
Hendy M. D., Penny D. Spectral analysis of phylogenetic data. J. Classif. (1993) 10:5–24.[CrossRef]
Hillis D. M. Taxonomic sampling, phylogenetic accuracy, and investigator bias. Syst. Biol. (1998) 47:3–8.
Hillis D. M., Huelsenbeck J. P. Signal, noise, and reliability in molecular phylogenetic analyses. J. Hered. (1992) 83:189–195.
Holland B. R., Delsuc F., Moulton V. Visualizing conflicting evolutionary hypotheses in large collections of trees: Using consensus networks to study the origins of placentals and hexapods. Syst. Biol. (2005) 54:66–76.
Holland B., Huber K., Dress A., Moulton V. Delta-plots: A tool for analyzing phylogenetic distance data. Mol. Biol. Evol. (2002) 19:2051–2059.
Holland B. R., Moulton V. Consensus networks: A method for visualising incompatibilities in collections of trees. In: WABI 2003 (2003) Berlin, Heidelberg, Budapest: Springer-Verlag. Pages 165–176.
Holland B. R., Penny D., Hendy M. D. Outgroup misplacement and phylogenetic inaccuracy under a molecular clock—a simulation study. Syst. Biol. (2003) 52:229–238.
Huber K. T., Langton M., Penny D., Moulton V., Hendy M. Spectronet: A package for computing spectra and median networks. Appl. Bioinformat. (2002) 1:159–161.
Huelsenbeck J. P. Is the Felsenstein zone a fly trap? Syst. Biol. (1997) 46:69–74.
Huelsenbeck J. P. Systematic bias in phylogenetic analysis: Is the Strepsiptera problem solved? Syst. Biol. (1998) 47:519–537.[Web of Science][Medline]
Johnsgard P. A. Cormorants, darters and pelicans of the world (1993) Washington, DC: Smithsonian Institution Press.
Kennedy M., Gray R. D., Spencer H. G. The phylogenetic relationships of the shags and cormorants: Can sequence data resolve a disagreement between behavior and morphology? Mol. Phylogenet. Evol. (2000) 17:345–359.[CrossRef]
Kennedy M., Spencer H. G. Phylogeny, biogeography, and taxonomy of Australasian teals. Auk (2000) 117:154–163.[CrossRef][Web of Science]
Kennedy M., Spencer H. G. Phylogenies of the Frigatebirds (Fregatidae) and Tropicbirds (Phaethonidae), two divergent groups of the traditional order Pelecaniformes, inferred from mitochondrial DNA sequences. Mol. Phylogenet. Evol. (2004) 31:31–38.[CrossRef][Web of Science][Medline]
Kennedy M., Spencer H. G., Gray R. D. Hop, step and gape: Do the social displays of the Pelecaniformes reflect phylogeny? Anim. Behav. (1996) 51:273–291.
Kocher T. D., Thomas W. K., Meyer A., Edwards S. V., Paabo S., Villablanca F. X., Wilson A. C. Dynamics of mitochondrial DNA evolution in animals: Amplification and sequencing with conserved primers. Proc. Natl. Acad. Sci. USA (1989) 86:6196–6200.
Larson A., Losos J. B. Phylogenetic systematics of adaptation. In: Adaptation—Rose M. R., Lauder G. V., eds. (1996) San Diego, California: Academic Press. Pages 187–220.
Lento G. M., Hickson R. E., Chambers G. K., Penny D. Use of spectral analysis to test hypotheses on the origin of pinnipeds. Mol. Biol. Evol. (1995) 12:28–52.[Abstract]
Lin Y. H., McLenachan P. A., Gore A. R., Phillips M. J., Ota R., Hendy M. D., Penny D. Four new mitochondrial genomes and the increased stability of evolutionary trees of mammals from improved taxon sampling. Mol. Biol. Evol. (2002) 19:2060–2070.
Lockhart P. J., Penny D., Meyer A. Testing the phylogeny of swordtail fishes using split decomposition and spectral analysis. J. Mol. Evol. (1995) 41:666–674.[Web of Science]
Lyons-Weiler J., Hoelzer G. A. Escaping from the Felsenstein Zone by detecting long branches in phylogenetic data. Mol. Phylogenet. Evol. (1997) 8:375–384.[CrossRef][Web of Science][Medline]
Mindell D. P., Sorenson M. D., Dimcheff D. E., Hasegawa M., Ast J. C., Yuri T. Interordinal relationships of birds and other reptiles based on whole mitochondrial genomes. Syst. Biol. (1999) 48:138–152.
Nelson G., Platnick N. Systematics and biogeography: Cladistics and vicariance (1981) New York: Columbia University Press.
Page R. D. M., Lee P. L. M., Becher S. A., Griffiths R., Clayton D. H. A different tempo of mitochondrial DNA evolution in birds and their parasitic lice. Mol. Phylogenet. Evol. (1998) 9:276–293.[CrossRef][Web of Science][Medline]
Paterson A. M., Gray R. D. Analytical tests of host/parasite cospeciation. In: Coevolutionary biology of birds and parasites—Clayton D. H., Moore J., eds. (1996) Oxford, UK: Oxford University Press. Pages 236–259.
Penny D., Hasegawa M., Waddell P. J., Hendy M. D. Mammalian evolution: Timing and implications from using the LogDeterminant transformation for proteins of differing amino acid composition. Syst. Biol. (1999) 48:76–93.
Poe S. The effect of taxonomic sampling on accuracy of phylogeny estimation: Test case of a known phylogeny. Mol. Biol. Evol. (1998) 15:1086–1090.[Web of Science]
Posada D., Crandall K. A. Modeltest: Testing the model of DNA substitution. Bioinformatics (1998) 14:817–818.
Rambaut A., Grassly N. C. Seq-Gen: An application for the Monte Carlo simulation of DNA sequence evolution along phylogenetic trees. Computer Applications in the Biosciences (1997) 13:235–238.
Ronquist F., Huelsenbeck J. P. MrBayes 3: Bayesian phylogenetic inference under mixed models. Bioinformatics (2003) 19:1572–1574.
Saitou N., Nei N. The neighbor-joining method: A new method for reconstruction of phylogenetic trees. Mol. Biol. Evol. (1987) 4:406–425.[Abstract]
Schulmeister S. Inconsistency of maximum parsimony revisited. Syst. Biol. (2004) 53:521–528.
Semple C., Steel M. Phylogenetics (2003) Oxford, UK: Oxford University Press.
Sibley C. G., Ahlquist J. E. Phylogeny and classification of birds: A study in molecular evolution (1990) New Haven, Connecticut: Yale University Press.
Siegel-Causey D. Phylogeny of the Pelecaniformes: Molecular systematics of a privative group. In: Avian molecular evolution and systematics—Mindell D. P., ed. (1997) San Diego, California: Academic Press. Pages 159–171.
Slikas B. Recognizing and testing homology of courtship displays in storks (Aves: Ciconiiformes: Ciconiidae). Evolution (1998) 52:884–893.[CrossRef][Web of Science]
Swofford D. L. PAUP*. Phylogenetic analysis using parsimony (*and other methods) (2002) Sunderland, Massachusetts: Sinauer Associates. Version 4.
Swofford D. L., Olsen 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 Associates. Pages 407–514.
Thorley J. L., Wilkinson M. Testing the phylogenetic stability of early tetrapods. J. Theor. Biol. (1999) 200:343–344.[CrossRef][Web of Science][Medline]
van Tuinen M., Butvill D. B., Kirsch J. A. W., Hedges S. B. Convergence and divergence in the evolution of aquatic birds. Proc. R. Soc. B (2001) 268:1345–1350.[Medline]
Wiens J. J., Hollingsworth B. D. War of the iguanas: Conflicting molecular and morphological phylogenies and long-branch attraction in iguanid lizards. Syst. Biol. (2000) 49:143–159.
Wilkinson M. Majority-rule reduced consensus trees and their use in bootstrapping. Mol. Biol. Evol. (1996) 13:437–444.[Abstract]
Wilkinson M., Lapointe F.-J., Gower D. J. Branch lengths and support. Syst. Biol. (2003) 52:127–130.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  W. White, S. Hills, R Gaddam, B. Holland, and D. Penny Treeness Triangles: Visualizing the Loss of Phylogenetic Signal Mol. Biol. Evol., September 1, 2007; 24(9): 2029 - 2039. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
C. Bleidorn, I. Eeckhaut, L. Podsiadlowski, N. Schult, D. McHugh, K. M. Halanych, M. C. Milinkovitch, and R. Tiedemann Mitochondrial Genome and Nuclear Sequence Data Support Myzostomida As Part of the Annelid Radiation Mol. Biol. Evol., August 1, 2007; 24(8): 1690 - 1701. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. Hughes, M. Kennedy, K. P. Johnson, R. L. Palma, and R. D. M. Page Multiple Cophylogenetic Analyses Reveal Frequent Cospeciation between Pelecaniform Birds and Pectinopygus Lice Syst Biol, April 1, 2007; 56(2): 232 - 251. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||