A method for inferring the rate of evolution of homologous characters that can potentially improve phylogenetic inference, resolve deep divergence and correct systematic biases.

Abstract

Current phylogenetic methods attempt to account for evolutionary rate variation across characters in a matrix. This is generally achieved by the use of sophisticated evolutionary models, combined with dense sampling of large numbers of characters. However, systematic biases and superimposed substitutions make this task very difficult. Model adequacy can sometimes be achieved at the cost of adding large numbers of free parameters, with each parameter being optimized according to some criterion, resulting in increased computation times and large variances in the model estimates. In this study, we develop a simple approach that estimates the relative evolutionary rate of each homologous character. The method that we describe uses the similarity between characters as a proxy for evolutionary rate. In this article, we work on the premise that if the character-state distribution of a homologous character is similar to many other characters, then this character is likely to be relatively slowly evolving. If the character-state distribution of a homologous character is not similar to many or any of the rest of the characters in a data set, then it is likely to be the result of rapid evolution. We show that in some test cases, at least, the premise can hold and the inferences are robust. Importantly, the method does not use a "starting tree" to make the inference and therefore is tree independent. We demonstrate that this approach can work as well as a maximum likelihood (ML) approach, though the ML method needs to have a known phylogeny, or at least a very good estimate of that phylogeny. We then demonstrate some uses for this method of analysis, including the improvement in phylogeny reconstruction for both deep-level and recent relationships and overcoming systematic biases such as base composition bias. Furthermore, we compare this approach to two well-established methods for reweighting or removing characters. These other methods are tree-based and we show that they can be systematically biased. We feel this method can be useful for phylogeny reconstruction, understanding evolutionary rate variation, and for understanding selection variation on different characters.

Full text

Page 1

Syst. Biol. 60(6):833–844, 2011
c ? The Author(s) 2011. Published by Oxford University Press, on behalf of the Society of Systematic Biologists. All rights reserved.
For Permissions, please email: journals.permissions@oup.com
DOI:10.1093/sysbio/syr064
Advance Access publication on July 29, 2011
A Method for Inferring the Rate of Evolution of Homologous Characters that Can
Potentially Improve Phylogenetic Inference, Resolve Deep Divergence and Correct
Systematic Biases
CARLA A. CUMMINS AND JAMES O. MCINERNEY∗
Molecular Evolution and Bioinformatics Unit, Department of Biology, National University of Ireland, Maynooth, Co. Kildare, Ireland;
∗Correspondence to be sent to: Molecular Evolution and Bioinformatics Unit, Department of Biology, National University of Ireland,
Maynooth, Co. Kildare, Ireland; E-mail: james.o.mcinerney@nuim.ie.
Received 3 September 2010; reviews returned 14 February 2011; accepted 2 March 2011
Associate Editor: Michael Charleston
Abstract.—Current phylogenetic methods attempt to account for evolutionary rate variation across characters in a ma-
trix. This is generally achieved by the use of sophisticated evolutionary models, combined with dense sampling of large
numbers of characters. However, systematic biases and superimposed substitutions make this task very difficult. Model
adequacy can sometimes be achieved at the cost of adding large numbers of free parameters, with each parameter being
optimized according to some criterion, resulting in increased computation times and large variances in the model estimates.
In this study, we develop a simple approach that estimates the relative evolutionary rate of each homologous character. The
method that we describe uses the similarity between characters as a proxy for evolutionary rate. In this article, we work
on the premise that if the character-state distribution of a homologous character is similar to many other characters, then
this character is likely to be relatively slowly evolving. If the character-state distribution of a homologous character is not
similar to many or any of the rest of the characters in a data set, then it is likely to be the result of rapid evolution. We
show that in some test cases, at least, the premise can hold and the inferences are robust. Importantly, the method does
not use a “starting tree” to make the inference and therefore is tree independent. We demonstrate that this approach can
work as well as a maximum likelihood (ML) approach, though the ML method needs to have a known phylogeny, or at
least a very good estimate of that phylogeny. We then demonstrate some uses for this method of analysis, including the
improvement in phylogeny reconstruction for both deep-level and recent relationships and overcoming systematic biases
such as base composition bias. Furthermore, we compare this approach to two well-established methods for reweighting or
removing characters. These other methods are tree-based and we show that they can be systematically biased. We feel this
method can be useful for phylogeny reconstruction, understanding evolutionary rate variation, and for understanding se-
lection variation on different characters. [Compatibility; maximum likelihood; maximum parsimony; molecular phylogeny
reconstruction; site rate variation; site removal; simulation; systematic bias.]
Homologous characters evolve at different rates.
Within a given data matrix, some characters might
evolve at an appropriate rate to resolve the branching
order of the taxa in question (Townsend 2007) whereas
others might exhibit high levels of homoplastic noise.
Some might be too slowly evolving and therefore mute
with respect to phylogenetic statements (Kluge and
Farris 1969; Delsuc et al. 2005; Philippe et al. 2005;
Townsend 2007). A character could be considered im-
portant if it contains useful information about the phy-
logeny of the group of interest and if it is relatively free
of homoplasy for that group. Therefore, for deep phylo-
genetic relationships, a slowly evolving character might
prove useful, whereas for shallower relationships, a
more rapidly evolving character could prove to be more
useful. Character-state substitution rate (i.e., the rate at
which a characters state is transformed into a different
state) is an important factor to consider when ranking
the informativeness of characters. Knowing a priori the
rate of evolution of a character can greatly facilitate the
treatment of characters for phylogeny reconstruction.
A number of efforts have been made to evaluate
character-specific evolutionary rates. Farris (1969) intro-
duced successive approximations character weighting
(SACW) in order to weight characters according to a
perceived importance assigned to them. This weighting
scheme sought to ensure that characters with a higher
degree of correlation with the phylogenetic history were
more highly regarded during reconstructions. Farris de-
fined this correlation as the consistency index (CI) for a
matrix, or the goodness of the fit of the characters within
the matrix to a given tree. The CI for an individual char-
acter on a particular tree is derived as the minimum
possible character length divided by the observed char-
acter length on the considered tree. So, when a character
fits on a tree without apparent homoplasy, the CI value
is unity. If additional ad hoc hypotheses need to be in-
voked to explain the evolution of the character on the
tree, then the CI value will be less than one (Farris 1969).
The CI for a data matrix is obtained by averaging the CI
values for all the characters in the matrix. Therefore, a
tree must be initially inferred. In his description of the
method, Farris preweighted characters according to a
weighting system devised by Le Quesne (1969), though
he indicated that initial character weights set to unity
would also work. As a consequence of the approach,
characters that tend to disagree with the initial tree are
given a lower weighting in subsequent analyses, in con-
trast to characters that tend to agree with this initial tree,
whose weight remains high.
In the late 1980s, Olsen (1987) noted that among-site
rate variation (ASRV) could cause problems in phyloge-
netic inference, and he attempted to accommodate this
variation using a model-based approach that employed
833
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 2

834
SYSTEMATIC BIOLOGY
VOL. 60
a normal distribution. Using a model to account for
rate variation across sites can increase the probability
of finding the correct phylogenetic tree topology com-
pared with a method that does not account for rate
variation (Yang 1993). By using an evolutionary model
that neglects to account for ASRV, sequences will ap-
pear to have undergone fewer mutations overall and
will appear to be more similar to their relatives com-
pared with an analysis using a model that accounts for
ASRV. Therefore, much of the effort to improve phy-
logeny reconstruction accuracy has focused on methods
that deal with accommodating site rate heterogeneous
data (Farris 1969; Yang 1996; Brinkmann and Philippe
1999; Hirt et al. 1999; Schmidt et al. 2002).
Yang (1996) modeled ASRV using the gamma distri-
bution. This distribution has some attractive properties,
particularly given that its shape can change from being
L shaped to being hill shaped, depending on the charac-
teristics of the alignment. Again, this approach tries to
incorporate rate variation and it assumes that site rate
heterogeneityiswellapproximatedbythismodel.How-
ever, assuming that all sites are free to vary will lead to
incorrect estimations when there are sites in the data set
that do not or cannot change (Yang 1996). In 1970, Fitch
and Markowitz (1970) proposed that for a protein there
might be two classes of sites—invariable and variable
and they suggested a method of analyzing molecular
alignments in order to determine how many positions
were invariable and how many were variable. These
invariable sites can also confound phylogeny recon-
struction and accentuate rate variation across sites. To
overcome these issues, some studies have experimented
with the removal of sites that violate assumptions of
the models that are being used. This has the effect of
reducing the range of site-to-site rate variation in the
data set.
As an example of a study that effectively reduced site-
to-site rate variation, Hirt et al. (1999) not only removed
invariant sites, but also removed sites they considered
to be fast evolving (Hirt et al. 1999). They identified
fast-evolving sites by using two different phylogenetic
trees and only removing sites that were considered to be
fast evolving on both topologies. In this case, removal
of both slow- and fast-evolving sites vastly improved
the support values for internal branches on the phylo-
genetic trees and resulted in a robust placement of the
Microsporidia.
Many different methods exist for the identification of
sites with a high substitution rate (Farris 1969; Kuhner
and Felsenstein 1994; Brinkmann and Philippe 1999;
Hansmann and Martin 2000; Schmidt et al. 2002; Pisani
2004). The majority, though not all, of these methods
are tree based. Tree based methods identify rapidly
evolving sites based on a tree either provided by the
user or inferred by the method before site identifica-
tion. For instance, TREE-PUZZLE (Schmidt et al. 2002)
and DNArates (Maidak et al. 1996; Olsen et al. 1998)
estimate evolutionary rates for each character based
on a given tree and process of character-state substi-
tution. TREE-PUZZLE can employ a discrete gamma
distribution to estimate site rates, with sites allocated
to a different category based on their likelihood score
on the tree. The DNArates program has been used
in conjunction with the fastDNAml program (Olsen
et al. 1994) in order to partition alignments of homolo-
gous characters into rate categories (Fischer and Palmer
2005). Fischer and Palmer (2005) used a procedure that
is not unlike the SACW approach in order to reweight
characters for subsequent analyses. For a data set that
was aimed at settling the placement of Microsporidia,
they found that early unweighted data sets resulted in
a variety of placements of the taxon, whereas succes-
sive rounds of character reweighting tended to result in
fewer tree topologies and finally the authors settled on a
placement of the microsporidia with the fungi that was
best supported by the successively reweighted data.
Brinkmann and Philippe (1999) developed a method
known as “slow-fast” where an alignment is split into
groups (Brinkmann and Philippe 1999; Kostka et al.
2008). These groups are generally user-defined taxo-
nomic groups. The evolutionary rate at a given site is
calculated as the sum of the number of changes at the
same position in all the groups individually. Although
groups are, technically, user defined, any prior knowl-
edge of the group will be based on previous tree infer-
ences and, therefore, the slow-fast method is, by proxy,
a tree-based method. In addition, due to the nature of
this method, it is not suitable for small data sets.
The problem with tree-based methods is that the true
tree is rarely known with certainty. Therefore use of an
incorrect initial tree can result in incorrect assignation
of an evolutionary rate to a character. Each character is
compared with the given tree topology, whether correct
or incorrect. A character is considered rapidly evolving
if it conflicts with the initial tree or has a high level of
homoplasy when mapped onto the tree. By assuming a
topology prior to site rate identification, a slowly evolv-
ing site could potentially appear to be rapidly evolving,
simply because the tree onto which it is mapped is in-
correct. This initial error can become a source for sys-
tematic biases. Therefore, it may be preferable to have
a method of determining evolutionary rate for a char-
acter that is independent of any a priori tree estimation
procedure.
Tree-independentapproachestodifferentiallyweight-
ing characters for phylogeny reconstruction include the
Le Quesne (1969) test of character compatibility, which
provided a “coefficient of character-state randomness”
that could be used, if desired, to exclude characters from
subsequent analysis. Essentially, this test evaluates two
characters and if they can be mapped onto the same tree
without homoplasy, then they are compatible, otherwise
they are incompatible. Characters that have the highest
amounts of incompatibilities with the other characters
might be considered candidates for removal prior to
subsequent phylogenetic analysis. Le Quesne (1989)
later introduced the notion of compatibility within data
being indicative of the level of phylogenetic informa-
tion. This work was further extended by Meacham
(1994), who developed his “Frequency of Compatibility
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 3

2011
CUMMINS AND McINERNEY—TREE-INDEPENDENT GENERATION OF EVOLUTIONARY RATES
835
Attainment” statistic. Wilkinson (1998) highlighted the
advantages of creating split patterns for sites when
detecting conflict. Conflict, as defined by Le Quesne,
becomes much easier to identify and rank when using a
universal coding system for sites. Pisani (2004) utilized
this idea to identify fast-evolving sites. According to
the method of Pisani (2004), each site in the alignment
receives an Le Quesne Probability (LQP) score, which
is “[...] the probability of a random character having
as low or lower incompatibility with the rest of the
data than does the original character.”. Pisani used this
probability measure to explore arthropod relationships
using different strategies for removal of characters with
differing LQP values.
Hansmann and Martin (2000), in contrast with the
compatibility strategies, proposed a very simplistic
non–tree-based method for identifying rapidly evolving
characters. They used the number of different character
states in an alignment column as a proxy for evolution-
ary rate (Hansmann and Martin 2000). They cite the in-
tuitiveness of the relationship between higher numbers
of polymorphisms at a site and speed of evolution at
that site. The set of most polymorphic characters would,
therefore, be enriched in homoplastic sites (Hansmann
and Martin 2000). However, each site is treated as a
separate entity and consequently, this approach does
not include information that may be contained in the
data set as a whole, apart from ranking the sites from
least to most polymorphic. In this paper, we present our
method, TIGER (Tree Independent Generation of Evo-
lutionary Rates), which is based on a similar concept to
Le Quesne (1989), Wilkinson (1998) and Pisani (2004).
TIGER analyzes similarity within characters (Wilkinson
1998). We expect that fast-evolving characters have lost
some, most, or all of their phylogenetic signal and there-
fore should demonstrate reduced similarity with other
sites that are more slowly evolving. Rather than com-
paring sites and only allowing them to be compatible or
incompatible, our method allows sites to be scored ac-
cording to varying degrees of similarity. This approach
should provide a more fine-grained or nuanced result
than the one that scores sites as being either compatible
or incompatible.
In this report, we analyze synthetic data sets in or-
der to explore the behavior of our approach and then,
to demonstrate the utility of the method, we analyze
two well-known problematic data sets. Additionally,
we show that tree-based site removal approaches have
significant problems, particularly when the data set
contains a systematic bias (e.g., convergent base compo-
sitional bias), whereas our tree-independent approach
can overcome these biases.
METHODS
Set Partitions
Our method is based on the analysis of set partitions
at each position in a matrix. This matrix could be any
type of data, including alignments of DNA or protein
sequences or a matrix of homologous morphological
characters.
A partition of a set X is a set of nonempty subsets of
X such that every element x in X is in exactly one of
these subsets. We treat each character in the matrix as
a set and partition this set based on character states. A
set partition is denoted, for example, as {{1}, {2, 3}, {4},
{5}} or 1/2,3/4/5. The partition 1/2,3/4/5 shows that
for this character, taxa 2 and 3 have the same character
statewhichisdifferentfromalltheothers, taxon1, taxon
4 and taxon 5 each have unique character states—both
different from each other and different from taxa 2 and
3. In this way, each character’s partition is determined
in order to enable pairwise comparisons with the rest
of the characters in the data set. For example, in a nu-
cleotide alignment of six taxa, character J = AAGGGC
and character K = TTCCCA (assuming the order of the
taxa is the same for both characters in this example). The
partition set for both J and K is 1,2/3,4,5/6, despite hav-
ing different character states.
Using this kind of data transformation, we can mea-
sure the degree of similarity between characters based
on the similarity of their set partitions. We find that a
character with a set partition that is similar to many
other characters in the data matrix can usually, though
not always, be a more slowly evolving character than a
character with a set partition that is less similar to the
rest of the characters in the matrix. Therefore, we can
use the average similarity of a character’s set partition
to the rest of the matrix as a proxy for evolutionary
rate.
The rate rifor the character at position i is defined as:
?
where n is the total number of characters in the matrix
and pa(i,j) is the partition agreement score. This is de-
fined as
pa(i,j) =
where |P(j)| is the number of groups in the partition of
the jth character and a(x,P(i)) equals 1 if x ⊆ A for some
A ∈ P(i). P(i) may be defined as a partition in character i.
It is important to note that, given two sets A and B,
if A ⊆ B, it is not necessarily commutative and, often,
B ?⊂ A. In this case, pa(i,j) ?= pa(j,i). Also, because the
rate is based on averaging of combinations of 1 or 0 val-
ues, it will always have a range between 0 and 1. A con-
stant site, that is, a site with only one character state,
will have r = 1 given that the pa, will be one for every
comparison.
For example, consider two sites A = CTTAA and B =
AGGGG with partition sets 1/2,3/4,5 and 1/2,3,4,5,
respectively. pa(A,B) = 0.5 because, out of two par-
titions in B ({1} and {2,3,4,5}), only {1} ⊆ P(A). Given
that {2,3,4,5} is not a subset of any partition in A, a({1},
P(A)) = 1 and a({2,3,4,5}, P(A)) = 0 ∴ pa(A,B) = 0.5.
As mentioned, this calculation is not commutative, so
pa(B,A) ?= 0.5. pa(B,A) = 1 because all partitions in
A ⊆ P(B).
ri=
j?=ipa(i,j)
n−1
(1)
?
x∈P(j)a(x,P(i))
|P(j)|
,
(2)
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 4

836
SYSTEMATIC BIOLOGY
VOL. 60
This approach is designed to measure how much a
particular character tends to agree with the other char-
acters in the data. If a character shares partitions with
many other characters, then it is likely that they hold
similar information. This may be viewed as a signal in
the data. Conversely, a character whose set partition
greatly differs from the other signals in the data may
be thought of as noise. To put it another way, a rapidly
evolving character is likely to have sustained multiple
substitutions, some or all of whom might be superim-
posed on earlier substitutions, therefore, this character
is more likely to have a set partition that agrees less with
more slowly evolving characters.
It is reasonable to suggest that a character that shares
few partitions with the majority of other characters
could be considered rapidly evolving. On the other
hand, a slowly evolving character is more likely to share
partitions with, or at least have fewer that conflict with,
many other characters. The first assumption might not
hold true in a situation where all or most characters in
a matrix are rapidly evolving. It is most likely to hold
true when evolutionary rates are moderate and when
there is a gradient of evolutionary rates from slow to
fast. Note that the rate of evolution that is assigned to a
particular character is measured in arbitrary units and
will vary with the data matrix being used. It is not a
measure of substitutions per unit of time and indeed
there are no units associated with the rate. This method
can be used to analyze DNA, protein, morphological, or
other arbitrary homologous characters.
It should be noted that for the current analyses, we
did not attempt to deal with missing data. Missing data
can be a feature of both molecular and morphological
data sets, usually because a particular gene or mor-
phological character has not been sampled or found.
Missing data can be accommodated by an appropriate
pruning of the characters so that only character states
that have been observed are being compared.
Binning
Itisoftenusefulorconvenienttogroupsiteswithsim-
ilar evolutionary rates together and in our implemen-
tation of this method a range of rates can be divided
into a user specified number of partitions, or bins. Sites
are placed into bins depending on their rate value. The
slowest rate and the fastest rate are determined and bins
are constructed by splitting the rates into equal parti-
tions. In this paper, we have used a variety of binning
schemes, from 8 bins to 20 bins. In theory, any num-
ber of bins can be constructed, as long as the number
is less than or equal to the number of characters in the
matrix.
Data Simulations
In order to test the features of the method, we
generated a number of artificial nucleotide data sets,
using a phylogenetic tree and a prespecified model of
nucleotide substitution. In the first instance, we simply
wanted to know if data sets with different patterns of
ASRV would return different patterns when analyzed
using TIGER. Second, we wanted to see if removing
characters had a beneficial effect on the fit of the data
matrix to all possible trees or produced the desirable
effect of improving the fit of the data to “good” trees
while worsening the fit of the data to bad” trees. Our
third simulation experiment involved the evaluation
of whether or not the TIGER approach to character re-
moval would improve the likelihood of resolving deep
relationships.
In this report, we have used nucleotide data for
reasons of ease of interpretation and also because of
the ready availability of excellent computer software
(Rambaut and Grassly 1997) to generate the data; how-
ever, in principle we could have used protein, morpho-
logical, or any kind of multistate character matrices.
Varying gamma shapes.—Using Seq-Gen (Rambaut and
Grassly 1997), we simulated two data sets over the same
49-taxon tree (Maddison 2004) (the tree is available in
Supplementary Material, available from http://www.
sysbio.oxfordjournals.org/) and we employed a model
that used a discrete approximation to the gamma distri-
bution, with four categories of sites. In order to assess
whether or not the TIGER algorithm could detect differ-
ent patterns of ASRV, two different α values were used
in simulations—0.5 and 20.0 reflecting two different dis-
tribution shapesthe first is L shaped and the second is
hill shaped. Both alignments were 999 bp in length and
simulated under the JC model (Jukes and Cantor 1969).
We experimented with other models of sequence evolu-
tion and different tree shapes and numbers of taxa and
the results are essentially the same as presented here, so
we only present the results of the JC simulations on this
data set.
Changing fit of the data to all trees in treespace.—Removal
of homoplastic characters in a matrix should have the
effect of improving the fit of the data to the true tree
whereas worsening the fit of the matrix to trees that are
very different from the true tree. However, given that it
is possible to edit any tree to change its topology into
any other tree, if we perform any data modification it
will most likely influence the goodness-of-fit of the data
to all trees in some way. Some trees are very similar
to the true tree and some are very dissimilar, conse-
quently, whereas incrementally removing larger num-
bers of characters (grouped into bins), we investigated
the change in fit of the data to all possible phylogenetic
trees for an eight-taxon data set. In our experiments, we
measured the change in the CI for all trees as bins were
sequentially removed, starting with the bin containing
the most rapidly evolving characters (a total of 10 bins
were used in this experiment). In effect, for the set of
all trees, T, we computed the CI for the original data
set on tree t (t ∈ T) and compared this value with the
CI value for the data set with Bin10 removed. We then
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 5

2011
CUMMINS AND McINERNEY—TREE-INDEPENDENT GENERATION OF EVOLUTIONARY RATES
837
plotted this value against the “nodal” distance (Puigbo
et al. 2007) between the true tree and tree t (when t
is not the true tree). For the true tree, the nodal dis-
tance is always zero. We carried out the same procedure
when we removed Bin9+Bin10, Bin8+Bin9+Bin10, and
Bin7+Bin8+Bin9+Bin10.
TIGER rates versus likelihood scores.—Using the correct
tree and the correct model, site-specific likelihood scores
can give a very good estimate of character evolutionary
rate. We wished to test how well the TIGER approach
could identify these characters without any knowledge
of a tree. We used 100 different seven-taxon trees chosen
at random from treespace (which contains 945 unrooted
trees). A nucleotide alignment of 999 positions was gen-
erated under the JC model for each of these 100 trees.
We generated site-specific likelihood scores in PAUP*
(Wilgenbusch and Swofford 2003) for all 945 trees for
each data set and we measured the ranking of sites on
each tree to TIGER rankings. That is to say, the site(s)
with the highest likelihood value are ranked as #1 and
the site(s) with the lowest value as #999 and likewise for
TIGER rates. The Euclidian distance between all likeli-
hood rankings and TIGER rankings was calculated. This
is a very simple measure of the average difference in
rank for a character in the two lists.
Deep branching tree.—Rapid evolution can obfuscate
deep relationships on a tree, often leading to unwanted
polytomies. This situation is particularly problematic
when long unbroken branches subtend a series of rapid
cladogenetic events. To test whether the TIGER ap-
proach could help resolve deep relationships where
there is very little phylogenetic signal, we used the JC
model of sequence evolution to produce 100 simulated
999 bp nucleotide data sets across the eight taxon tree
shown in Figure 2. The short deep branches combined
with long terminal branches presents a difficult problem
for phylogenetic analysis, mostly due to the confound-
ing effects of rapidly evolving characters. To ensure
that the data generated displayed poor phylogenetic
resolution, we built a majority-rule consensus tree from
maximum likelihood (ML) trees constructed from each
of the data sets prior to any site removal. This was re-
peated after removal of sites dictated by TIGER and
to test the performance of a tree based method in this
scenario, we also repeated the analysis after removal of
rapidly evolving sites identified by ML. The ML tree
was estimated using PAUP* and the sites were catego-
rized on this tree using TREE-PUZZLE.
Empirical Testing
Thermus data set.—In order to further understand
TIGER’s functionality, two empirical data sets were
used. A 1273-column alignment of bacterial 16S ribo-
somal RNA genes known as the Thermus data set is
well studied (Embley et al. 1993; Mooers and Holmes
2000), and we used this data set to examine whether
the TIGER approach is useful for accounting for base
compositional biases. This data set contains three ther-
mophiles, Aquifex aeolicus, Thermatoga maritima, and
Thermus aquaticus whose sequences are enriched in G
and C nucleotides and two mesophiles, Bacillus subtilis
and Deinococcus radiodurans whose nucleotide composi-
tion is more balanced. A combination of compositional
bias and distant relationships can mean that when there
is only a weak phylogenetic signal, it can be overcome
by the similarity in base composition of the most rapidly
evolving positions in the alignment. In general, many
methods of phylogenetic analysis will group the ther-
mophiles together in this data set, despite the fact that
there is strong evidence that T. aquaticus and D. radio-
durans are sister taxa (Embley et al. 1993). We refer to a
tree displaying the mesophiles as a monophyletic group
to the exclusion of the thermophiles as the ATTRACT
tree and this is the tree recovered by most tree infer-
ence methods using the whole sequence alignment. We
refer to a phylogenetic tree that places T. aquaticus and
D. radiodurans together as the TRUE tree. Due to this
well-characterized strong compositional attraction, we
wished to investigate whether site removal using the
TIGER approach could influence recovery of the correct
tree.However,todemonstratethedifferenteffectsofsite
removal in a tree-independent fashion compared with
the traditional ML approaches, we also compared the
topology inferred after removal of rapidly evolving sites
identified by TIGER with the topology recovered after
removal of rapidly evolving sites according to TREE-
PUZZLE (Schmidt et al. 2002) and SACW (Farris 1969).
We did not use TREE-PUZZLE to infer the tree, we sim-
ply used the method implemented by TREE-PUZZLE
to assign evolutionary rates to sites, based on a tree that
we supplied to the software.
Primate data set.—It has generally been accepted that
humans share a close relationship with orangutans, go-
rillas, and chimpanzees (Hayasaka et al. 1988; Begun
1992; Adachi and Hasegawa 1995; Shoshani et al. 1996;
Ruvolo 1997; Satta et al. 2000; Ebersberger et al. 2007).
From this group, it is generally agreed that orangutans
are the least closely related to humans and that hu-
mans, chimps, and gorillas form a monophyletic group,
though there are some conflicting opinions (Schwartz
1984; Grehan and Schwartz 2009).
The relationships of interest, therefore, concern the
human, chimpanzee, and gorilla lineages (Satta et al.
2000). The separation of these three lineages is thought
to have occurred in quick succession (Hayasaka et al.
1988; Adachi and Hasegawa 1995), and this makes the
phylogeny difficult to resolve and the two alternative
hypotheses—human, chimp together (HC hypothesis)
or chimp, gorilla together (CG hypothesis)—receive
almost equal support from this data set. Because of
the controversy surrounding this topology, the second
empirical data set we use is a well-known primate
mitochondrial data set (see Supplementary material)
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 6

838
SYSTEMATIC BIOLOGY
VOL. 60
consisting of 12 sequences and 898 aligned nucleotide
positions (Hayasaka et al. 1988).
In a parsimony analysis of the data set, with all char-
acters being equally weighted, both the HC and the CG
hypotheses are equally good, with 1153 steps required
to explain the data. We used the tree-based methods
of assigning character evolutionary rates and use al-
ternatively the HC and the CG trees in order to carry
out the inferences. We compared and contrasted the re-
sults from tree-based analysis with the tree-independent
method described here.
RESULTS AND DISCUSSION
Varying Gamma Shapes
OurfirstanalysisofthebehavioroftheTIGERmethod
focused on the analysis of simulated data sets for 49 taxa
with different patterns of rate variation across sites. We
chose the 49-taxon data set that is distributed with the
MACCLADE software (Maddison 2004) because it con-
tains a reasonable range of branch lengths and has a
moderately large number of taxa. We simulated two
separate data sets that differed by the ASRV model used
to generate the data. In the first case, we used a gamma
distribution with an α parameter of 20 and in the second
the α parameter was set to 0.5, reflecting very different
evolutionary scenarios. We then used the TIGER ap-
proach to place sites into 20 bins sorted by their rate of
evolution (Fig. 1a,b).
There are two interesting points to be made about Fig-
ure 1. First of all, the two graphs are not the same and
furthermore Figure 1b, which is generated from the data
set with an α parameter of 0.5, is more L shaped than
Figure 1a, which was generated from the data with an α
parameter of 20. This indicates that the TIGER approach
is detecting the different ASRV patterns. What is of fur-
ther interest is that within each graph there is a clear
multimodality. There are four clusters of bars on the his-
tograms (indicated by the alternative shading and clear
zones on the diagrams). When the seq-gen software
generates data, it uses an approximation to the gamma
distribution and in these cases an approximation was
employed that used four categories of sites. The TIGER
approach has identified these subtle patterns and has
placed the different sites into clusters.
True Tree versus Incorrect Trees
If the removal of rapidly evolving characters really
is a good idea for improving the chances of recovering
the correct phylogenetic tree, then we expect that re-
moval of these characters would improve the goodness-
of-fit of the data to the true tree while worsening the
goodness-of-fit of the data to other trees. In order to test
this hypothesis, we generated a simulated data set con-
taining eight taxa and using the JC model, according to
the protocols previously described. We progressively re-
movedthefastestevolvingsites,asjudgedbytheTIGER
approach, until we had removed the four fastest cate-
gories of sites. We then examined the goodness-of-fit of
the data to the correct tree (the tree used to simulate the
data) and also the goodness-of-fit of the data to all the
other possible trees. We plotted the goodness-of-fit mea-
sure (CI) against the nodal distance (as measured by the
TOPD/FMTS software, Puigbo et al. 2007) for the un-
stripped data set for each possible tree topology and we
plotted the change in CI (ΔCI) against nodal distance for
each of the data sets where sites were stripped. The re-
sults of these experiments are seen in Figure 2. In total,
there were 10,395 trees examined for each treatment of
the data.
With all sites included in the alignment, the CI for the
correct tree was 0.825. The worst CI value in the data set
was 0.612 and the tree with the largest nodal distance
from the true tree had a distance of 2.44949 and a CI
value of 0.616. In general, there is a negative correlation
between CI and nodal distance from the true tree.
When we stripped out the Bin10 category of sites,
we saw the CI values increased for some trees and de-
creased for others. The CI value with the largest increase
for any of the 10,395 trees was the CI value for the true
FIGURE 1.
tree in S1 and ASRV modeled using a gamma distribution with a shape parameter of 0.5, and (b) data set of the same size and topology but
with ASRV modeled using a gamma shape parameter of 20.0. The alternating shaded and clear areas indicate the four categories of sites that
approximate the gamma distribution. This figure is available in black and white in print and in color at Systematic Biology online.
Histograms of binning results for two different data sets with different ASRV. a) A 999-bp, 49-taxon data set generated using the
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 7

2011
CUMMINS AND McINERNEY—TREE-INDEPENDENT GENERATION OF EVOLUTIONARY RATES
839
FIGURE 2.
tree—an increase to 0.852. In contrast, the tree with the
largest nodal distance from the true tree experienced
a decrease in CI value and its new value was 0.612.
Overall, a total of 5364 trees (51.6% of the total) saw an
increase in CI value, whereas 5031 trees experienced a
decrease in CI value.
Continued site stripping resulted in a progressive in-
crease in CI value for the true tree and a progressive
decrease in CI value for the tree with the largest nodal
distance from the true tree. When Bin categories 9 and
10 were removed, the values changed to 0.894 and 0.609,
respectively, with 5403 (51.9%) of the trees now experi-
encing an increase in CI value. When Bin categories 8,
9 and 10 were removed, the values changed to 0.911 for
the true tree and 0.601 for the worst tree with 3811 of the
trees having an increased CI value. Finally, when we re-
moved Bin categories 7, 8, 9, and 10, the values changed
to 0.923 and 0.597, respectively, with 3257 of the trees ex-
periencing an increase in CI value (31.3%), whereas 7138
had a decreased CI value (68.6%).
Therefore, we can see for this data set that site strip-
ping has resulted in a bias in the fit of the data to differ-
ent trees. In general, those tree topologies that are close
to the true tree will begin to fit the data better, whereas
those trees that are least similar in topology to the true
will begin to fit the data worse. The tree that is most
positively affected by site stripping is the true tree. It
must be remembered that the TIGER approach is not
tree based and at no time was the TIGER software aware
of the topology of the true tree.
TIGER Rates versus Likelihood Scores
To see how well TIGER can approximate site-specific
rates we compared it with likelihood scores for each site
on every possible seven-taxon unrooted tree. The Eu-
clidian distance from TIGER ranking to the likelihood
rankings on all trees were recorded for all data sets,
with particular emphasis on where the distance be-
tween TIGER rankings and the likelihood rankings on
the known true tree fell with respect to the other trees.
In 100% of data sets, this distance fell within the top
0.3% of all scores. In 95% of all cases, the distance from
TIGER rankings to the likelihood rankings on the true
tree was the smallest distance recorded to any tree in
the data set.
This shows that the TIGER approach will produce an
ordering of the evolutionary rates of the sites that is usu-
ally closer to the ranking of sites according to the true
tree than to other incorrect trees.
Deep Branching Tree
In order to see whether it is possible for our method
to improve the resolution of deep relationships where
←
scissa is the nodal distance of a tree from the correct tree and on the
ordinate is either the CI or the difference in CI value between the un-
stripped alignment and the stripped alignment (ΔCI).
FIGURE 2.Change in CI with increasing site removal. On the ab-
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 8

840
SYSTEMATIC BIOLOGY
VOL. 60
phylogenetic signal is weak, we simulated 100 differ-
ent DNA alignments based upon a single phylogenetic
tree with long external branches and very short internal
branches (Fig. 3a). This alignment was designed to rep-
resent a difficult problem of phylogenetic inference and
was simulated using the JC model of sequence evolu-
tion. ML trees for each of the data sets were inferred
under the JC model. As expected, prior to removal of
rapidly evolving sites, the majority-rule consensus anal-
ysisusingtheJCmodelproduceda treewithpolytomies
and poor resolution (Fig. 3b), and the only branch that
is resolved has a bipartition frequency (BF) of 55% was
for a split that separates taxa C and D from the rest. We
used the TIGER approach to identify the rapidly evolv-
ing characters in the matrices and place all characters
into 10 bins with increasing evolutionary rate. Removal
of the most rapid category of sites, Bin10, which con-
tained between 183 and 502 sites with an average of
424 between the 100 data sets, entirely resolved all ex-
cept the basal polytomy (Fig. 3c), with BF ranging from
67% to 99%. We wished to test our method against a
FIGURE 3.
netic events. a) The topology of the tree used to generate the simulated
data (see text for details of simulation). b) Majority-rule consensus ML
tree after before site removal and also after site removal using ML.
The bootstrap support value for the unstripped alignments is above
the line and the value after site removal using likelihood is below the
line. c) Majority-rule consensus ML tree after removal of Bin10, the
fastest evolving sites, according to the TIGER method.
Effect of site removal on deep closely spaced cladoge-
tree-based method. We used TREE-PUZZLE (Schmidt
et al. 2002) on the same simulated data. Removing the
most rapidly evolving category of sites using the TREE-
PUZZLE approach (ranging from 269 to 481 sites, mean
of 334 sites removed) the tree remained equally unre-
solved as prior to any site removal, with the BF of the
split separating C and D rising to 61 (Fig. 3b).
This shows both the pitfall of the tree-based method
and the advantage of our tree-independent method.
The sites identified as most rapidly evolving by TREE-
PUZZLE are those that do not agree with the initial
tree inferred by ML. For this reason, removal of these
sites does not clarify signals in the data, rather it merely
strengthens the signal for the initial groupings. The
tree-independent method, however, does not need any
initial tree, therefore it is not biased toward any single
tree and, instead, it picks out genuine signals in the
data.
Thermus Data Set
The Thermus data set consists of 1273 aligned
nucleotide positions from the 16S rRNA gene and is
available as Supplementary Material. Using ML phylo-
genetic reconstruction implemented in PAUP4.0b10, we
examined the differences in tree topology when remov-
ing characters judged to be rapidly evolving according
to TIGER versus characters judged to be rapidly evolv-
ing according to TREE-PUZZLE (with a user-supplied
tree, constructed using ML). In addition, we used the
reweight command in PAUP to apply SACW (Farris
1969) and evaluate the effect that this approach had
on the chances of recovering the correct tree. Using
the original alignment of 1273 aligned positions (see
Supplementary Material) and a GTR+I+G model of se-
quence evolution, we produced the phylogenetic tree
in Figure 4a. Using the TREE-PUZZLE software, we cat-
egorized sites according to the GTR+I+G model using
a discrete approximation to the gamma distribution to
model ASRV, with a total of eight categories of sites.
The category of sites with the fastest rate of evolution
was removed from the alignment (a total of 186 sites)
and the analysis was re-run using this newer shorter
data set (consisting of 1087 sites). In this case, the same
ATTRACT tree was recovered. The most significant dif-
ference between the two bootstrap analyses was that the
bootstrap support values for the data set with the sites
removed were much higher and each of the internal
nodes was recovered in 100% of the bootstrap pseu-
doreplicates (Fig. 4b). It must be remembered that the
rates of evolution of the sites had been determined us-
ing the ATTRACT tree, which is the tree that is obtained
in the analysis of the unstripped data set.
In order to investigate the SACW method, we first in-
ferred the most parsimonious phylogenetic tree with all
sites equally weighted and using an exhaustive search
of tree space and the parsimony optimality criterion.
Support for this tree was assessed using 1000 rounds
of bootstrap resampling, with the results summarized
by a majority-rule consensus procedure. The most
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 9

2011
CUMMINS AND McINERNEY—TREE-INDEPENDENT GENERATION OF EVOLUTIONARY RATES
841
FIGURE 4.
port prior to site removal. b) The tree recovered after removal of sites
identified by PUZZLE and using SACW. c) The resulting tree after re-
moval of sites identified by TIGER.
Analysis of the Thermus data set. a) Topology and sup-
parsimonious tree was once again the ATTRACT tree,
with bootstrap support values of 92% for the grouping
of D. radiodurans and B. subtilis and 96% for a clan con-
taining A. aeolicus and T. maritima. Using the reweight
command in the PAUP software, we weighted the char-
acters according to their CI value on this tree. We then
carried out another bootstrap resampling analysis to
assess support for groups on the tree. This time the AT-
TRACT tree was once again recovered, but the support
for all internal edges was at 100%.
We used the TIGER approach to identify rapidly
evolving sites in the rRNA data set. We placed all sites
from the alignment into one of eight bins according to
how rapidly they evolve. The most rapidly evolving
category of sites contained 108 sites and these were re-
moved for subsequent ML analysis. Using the GTR+I+G
model of sequence evolution on the remaining 1165
sites, we recovered the TRUE phylogenetic tree. After
1000 bootstrap replicates, we observed that the group-
ing of D. radiodurans and T. aquaticus in 81% of the repli-
cates and the grouping of T. maritima and B. subtilis was
observed in 68% of the replicates. The ATTRACT topol-
ogy that groups D. radiodurans and B. subtilis together
was seen in 19% of the replicates.
We carried out an additional analysis of the sites that
are identified as being rapidly evolving. In all cases, we
analyzed the most rapidly evolving sites on their own
to see if there was any strong phylogenetic signal in
those sites. As these sites are saturated for change, we
do not expect to see a single phylogenetic signal, rather
a number of incongruent signals. In our analyses, only
the sites in Category 8 of the ML analysis contained any
congruent phylogenetic signal. There was 80% boot-
strap support for the TRUE tree in these sites. This result
demonstrates that not only does such an ML approach
result in strong support for the incorrect topology but
also the characters that it discards contain more true
phylogenetic signal than the characters that it retains.
This needs to be viewed as a systematic error.
Primate Data Set
Our last analysis involves an 898 bp data set of 12 pri-
mate mitochondrial sequences (Hayasaka et al. 1988).
Two equally most parsimonious trees, requiring 1153
steps can be obtained by analysis of these sequences.
One of these trees places the human and chimpanzee
(Pan troglodytes) together as sister taxa, whereas the
other tree groups the Chimpanzee with the Gorilla. We
wanted to investigate two things with this data set.
First, in this case, where two phylogenetic hypotheses
are strongly competing and where there is no greater
support for one topology over the other, whether the
TIGER approach would recover the accepted tree (hu-
man and chimp together) with confidence. Second,
whether the tree-dependent method would be influ-
enced strongly by the tree that is used to determine
the evolutionary rate of the characters, or whether it
would work well irrespective of the tree that it used
initially for character reweighting. More specifically,
we wished to see if using a particular tree in order to
generate evolutionary rates would tilt the balance in
favor of this topology in a bootstrap analysis. In other
words, we wanted to explore whether character re-
moval, based on an incorrect tree, could override the
(albeit small) amount of extra support for the true tree
and subsequently provide strong support for the incor-
rect tree.
When the tree that places Homo and Pan together
was used in SACW in order to reweight characters ac-
cording to the CI, then this same tree was recovered in
the majority-rule consensus tree following bootstrap-
ping. The bootstrap support value for this relationship
was 79%, compared with a 51% value for the equally
weighted data set (10,000 bootstrap replicates). We then
used the other equally parsimonious tree in order to
carry out character weighting for SACW. Using char-
acter reweighting according to the CI, we obtained a
bootstrap support value of 77% for the grouping of Pan
and Gorilla together. This shows that the initial tree that
is used for character weighting can override small phy-
logenetic signals and because characters that tend not to
agree with this initial tree are down weighted, this has
a huge affect on which tree is supported in subsequent
analyses.
It should be noted that in this particular case, the
ML approach to site stripping was not as sensitive as
the SACW approach and indeed was quite insensi-
tive to the initial tree that was used for site classifica-
tion. When the HC hypothesis tree was used, and the
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 10

842
SYSTEMATIC BIOLOGY
VOL. 60
TREE-PUZZLE software was asked to put sites into a
total of 10 categories, then a total of 114 sites were put
into the fastest category. When the CG hypothesis tree
was used, then a total of 121 sites were put into the
fastest category. Irrespective of the tree that was used
to categorize sites, when category 10 was removed, we
always recovered strong support for the HC hypothesis.
We should note, however, that when the HC hypothesis
was used to categorize sites, the resulting bootstrap sup-
port value was 99%, whereas when the CG hypothesis
tree was used to categorize sites, then support for the
HC hypothesis after site stripping was somewhat lower
at 81%.
We used the TIGER approach to categorize characters
in a tree-independent manner and to place them into a
total of 10 bins according to their average split similar-
ity with the other characters in the matrix. We removed
the fastest category of sites, Bin10, which contained a
total of 192 characters. We then used maximum parsi-
mony bootstrapping to evaluate support for groups in
the phylogeny. We recovered a grouping of Homo and
Pan, with 87% support after 10,000 bootstrap replicates.
The alternative hypothesis, grouping Pan and Gorilla
together received 8.8% bootstrap support. Using ML,
the HC hypothesis received 90% bootstrap support,
whereas the CG hypothesis received 6% support.
CONCLUSION
In this article, we report the development of an algo-
rithm, based on those of Le Quesne (1989), Wilkinson
(1998) and Pisani (2004) that uses similarity in the pat-
tern of character-state distributions between characters
as a proxy for speed of evolution in a data matrix of
homologous characters. We expect that rapidly evolv-
ing characters are likely to lose some, most, or all of
their phylogenetic information and will tend to have a
character-state distribution that is closer to random than
the distribution expected from a more slowly evolving
character. A character is assumed to be rapidly evolving
if it has a character-state distribution pattern that, on
average, is not very similar to the patterns observed in
other characters. This assumption is only likely to hold
in some (though probably very many) situations. Specif-
ically,inadatamatrixwhereeachcharacteriseffectively
randomized, due to a very rapid rate of evolution or a
long evolutionary timespan, we do not expect that this
kind of approach will work well. Notwithstanding this
caveat (which is a situation that would confound most,
if not all, phylogenetic methods), we have observed
some very interesting and desirable properties of this
approach that make it a useful addition to the phyloge-
netic arsenal.
The TIGER approach identified differing patterns
of ASRV, distinguishing alignments that had extreme
variation in among-site evolutionary rates from those
alignments that had a more even distribution of rates.
Additionally, it was able to identify subtleties in the data
such as the four clusters of rates in each alignment—a
by-product of the simulation process.
The TIGER approach helped improve the fit of the
data to the correct tree in our simulations. Removing
sites that TIGER identified as being rapidly evolving
resulted in a better fit of the data to good trees and
worse fit of the data to bad trees, with the true tree
being affected most positively. Additionally, using the
TIGER approach, we could improve the resolution of
deep lineages where rapid cladogenesis resulted in very
difficult-to-resolve branches. Worryingly, the likelihood
approach to removing rapidly evolving sites proved
to be problematic—the sites that were removed were
those that did not agree with the initial tree, resulting
in a situation where, out of 100 simulations, there was
little improvement in the recovery of the deep diverging
rapid cladogenesis tree.
For the ribosomal RNA data set, we observed a num-
ber of issues. First, the TIGER approach seems to have
some merit as an approach to removing sites that inter-
fere with phylogeny reconstruction. Additionally, two
other tree-dependent methods—site identification using
a ML model of ASRV and site identification using the
fit of the data to an initially constructed phylogenetic
tree—are systematically biased toward favoring the first
phylogenetic tree they construct. We, therefore, feel it is
important to be cautious when using tree-based meth-
ods of assigning evolutionary rates to sites, unless the
evolutionary history is known with certainty. We note,
however, that a sophisticated compositionally heteroge-
neous model of sequence evolution is capable of identi-
fying the correct topology for this data set, without the
necessity of deleting or reweighting characters (Foster
2004).
The point concerning tree-based attribution of evolu-
tionary rate is quite clearly exemplified by the primate
mitochondrial data set and maximum parsimony anal-
ysis. Here two hypotheses are equally good when using
the parsimony criterion. Character reweighting based
on one of the two equally most parsimonious trees
will skew subsequent analyses toward supporting this
particular topology, whereas the same is true for the
alternative topology. Ultimately, the TIGER analysis,
which does not use a tree, recovers the correct phylo-
genetic hypothesis (which has been confirmed by nu-
merous other studies) while not using an a priori deter-
mined phylogenetic tree in order to do so. We find that
support for the grouping of Pan and Gorilla, to the exclu-
sion of Homo is an artifact that is due to the most rapidly
evolving sites. This also shows that site stripping can
be beneficial for resolution of recent relationships, not
just ancient relationships. We should also state here that
ML analysis of this data set produces the correct tree,
using the Tamura–Nei model, with bootstrap support
for (Homo, Pan) at 94%.
Ultimately, TIGER is an interesting device for iden-
tifying characters that do not agree with the majority
of the data. We argue here that in many cases this dis-
agreement can be diagnostic of rapid evolution. At the
very least, the converse is likely to be true—rapid char-
acter evolution is likely to produce a pattern that is not
very similar to other characters. Removal of these kinds
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 11

2011
CUMMINS AND McINERNEY—TREE-INDEPENDENT GENERATION OF EVOLUTIONARY RATES
843
of characters can greatly improve the accuracy of suc-
cessive phylogenetic analysis by removing conflicting
signals.
There are surely limits to what site removal can
accomplish and with certainty site removal is a poor
alternative to precise model definition. However, pre-
cise model definition comes with a cost. Models that
adequately describe the evolution of a set of DNA or
protein sequences might, of necessity, be very parameter
rich (e.g., using a combination of Dirichlet processes for
both site rate identification, Huelsenbeck and Suchard
2007, and site-specific profiling, Lartillot and Philippe
2004, as implemented in the CAT model) and require
a large amount of sequence before they become sta-
tistically consistent. The most commonly used models
of sequence evolution are often inadequate to describe
the evolution of the sequences being studied. Model
selection approaches often “max-out,” where the most
parameter-rich method of analysis is the one that is
selected by a likelihood ratio test, Akaike information
criterion or Bayesian information criterion (Keane et al.
2006), indicating that perhaps there are not enough
parameters available. Therefore, it might not be an op-
tion to use a precisely described model. In the case of
the rRNA sequences being analysed in this study, the
raw alignment exhibited significant compositional het-
erogeneity and none of the standard, compositionally
homogeneous time-reversible models of sequence evo-
lution can adequately account for this heterogeneity.
By identifying and removing the most rapidly evolving
characters, the models are better able to account for the
evolution of the sequences.
We have no good theoretical framework for knowing
precisely how many sites to remove from an alignment.
It is likely that in many cases there is no need to strip
out any sites. At the moment, we only have an ad hoc
approach to site stripping and this must be considered
a major problem. Ideally, we wish to remove sites that
only contribute noise and do not contribute any phy-
logenetic signal. Our recommendation is the testing of
congruence across a progressively larger number of the
fastest evolving characters using methods such as the
permutation-tail-probability test (Faith and Cranston
1991) or likelihood mapping (Strimmer and von Hae-
seler 1997). This would result in the removal of sites that
show very little consistency with the rest of the data and
very little consistency with one another. However, this
is also ad hoc and should be used as nothing more than
a rule-of-thumb.
WealsonotethatbootstrapsupportvaluesorBayesian
clade probability values are probably meaningless when
there is a directed attempt to remove sites that disagree
with the rest of the data. It is likely that the support
values will tend to increase when incongruent data
are removed. When we use bootstrap support values,
we wish to show that the data have been strongly in-
fluenced by the character removal; we do not wish to
imply that bootstrapping should follow character re-
moval, as, in most cases, the resulting bootstrap scores
are likely to be higher.
Given that there are limits to what can be achieved
by character removal, we conclude by advising that this
method should be used as one part of an overall exper-
imental programme of data exploration. We expect that
additional tree-independent methods of analyzing evo-
lutionary rate variation can be developed.
SUPPLEMENTARY MATERIAL
Supplementarymaterialcanbefoundat http://www.
sysbio.oxfordjournals.org/.
FUNDING
C.A.C. is funded by a Science Foundation Ireland
ResearchFrontiersProgramme
EEEOBF654] to J.O.Mc.I.
award[07/RFP/
ACKNOWLEDGMENTS
The work was supported by the NUIM High-
performance computing resource and the Irish Centre
for High End Computing.
REFERENCES
Adachi J., Hasegawa M. 1995. Improved dating of the hu-
man/chimpanzee separation in the mitochondrial DNA tree: het-
erogeneity among amino acid sites. J. Mol. Evol. 40:622–628.
Begun D. 1992. Miocene fossil hominids and the chimp human clade.
Science. 257:1929–1933.
Brinkmann H., Philippe H. 1999. Archaea sister group of bacteria? In-
dications from tree reconstruction artifacts in ancient phylogenies.
Mol. Biol. Evol. 16:817–825.
Delsuc F., Brinkmann H., Philippe H. 2005. Phylogenomics and the
reconstruction of the tree of life. Nat. Rev. Genet. 6:361–375.
Ebersberger I., Galgoczy P., Taudien S., Taenzer S., Platzer M., von
Haeseler A. 2007. Mapping human genetic ancestry. Mol. Biol. Evol.
24:2266–2276.
Embley T., Thomas R., Williams R. 1993. Reduced thermophilic bias in
the 16S rDNA sequence from Thermus ruber provides further sup-
port for a relationship between Thermus and Deinococcus. Syst.
Appl. Microbiol. 16:25–29.
FaithD.,CranstonP.1991.Couldacladogramthisshorthavearisenby
chance alone?: on permutation tests for cladistic structure. Cladis-
tics. 7:1–28.
Farris J. 1969. Successive approximations approach to character
weighting. Syst. Zool. 18:374–385.
Fischer W. M., Palmer J. D. 2005. Evidence from small-subunit ribo-
somal RNA sequences for a fungal origin of Microsporidia. Mol.
Phylogenet. Evol. 36:606–622.
Fitch W., Markowitz E. 1970. An improved method for determining
codon variability in a gene and its application to the rate of fixation
of mutations in evolution. Biochem. Genet. 4:579–593.
Foster P. G. 2004. Modeling compositional heterogeneity. Syst. Biol.
53:485–495.
Grehan J., Schwartz J. 2009. Evolution of the second orangutan:
phylogeny and biogeography of hominid origins. J. Biogeogr. 36:
1823–1844.
Hansmann S., Martin W. 2000. Phylogeny of 33 ribosomal and six
other proteins encoded in an ancient gene cluster that is con-
served across prokaryotic genomes: influence of excluding poorly
alignable sites from analysis. Int. J. Syst. Evol. Microbiol. 50:1655–
1663.
Hayasaka K., Gojobori T., Horai S. 1988. Molecular phylogeny and
evolution of primate mitochondrial DNA. Mol. Biol. Evol. 5:
626–644.
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from

Page 12

844
SYSTEMATIC BIOLOGY
VOL. 60
Hirt R., Logsdon J., Healy B., Dorey M., Doolittle W., and Embley T.
1999. Microsporidia are related to fungi: evidence from the largest
subunit of RNA polymerase II and other proteins. Proc. Natl. Acad.
Sci. U.S.A. 96:580–585.
Huelsenbeck J., Suchard M. 2007. A nonparametric method for ac-
commodating and testing across-site rate variation. Syst. Biol. 56:
975–987.
Jukes T., Cantor C. 1969. Evolution of protein molecules. In: Munro,
editor. Mammalian protein metabolism. New York: Academic
Press. p. 240–253.
Keane T., Creevey C., Pentony M., Naughton T., McInerney J. 2006.
Assessment of methods for amino acid matrix selection and their
use on empirical data shows that ad hoc assumptions for choice of
matrix are not justified. BMC Evol. Biol. 6.
Kluge A. G., Farris J. S. 1969. Quantitative phyletics and the evolution
of anurans. Syst. Zool. 18:1–32.
Kostka M., Uzlikova M., Cepicka I., Flegr J. 2008. SlowFaster, a user-
friendly program for slow-fast analysis and its application on phy-
logeny of Blastocystis. BMC Bioinformatics. 9:341.
Kuhner M. K., Felsenstein J. 1994. A simulation comparison of phy-
logeny algorithms under equal and unequal evolutionary rates.
Mol. Biol. Evol. 11:459–468.
LartillotN.,PhilippeH.2004.Abayesianmixturemodelforacross-site
heterogeneities in the amino-acid replacement process. Mol. Biol.
Evol. 21:1095–1109.
Le Quesne W. 1969. A method of selection of characters in numerical
taxonomy. Syst. Biol. 18:201–205.
Le Quesne W. 1989. The normal deviate test of phylogenetic value of a
data matrix. Syst. Zool. 38:51–54.
MaddisonD.R.2004.Testingmonophylyofagroupofbeetles.Study1
in Mesquite: A Modular System for Evolutionary Analysis. Version
1.04. Available from: http://mesquiteproject.org.
Maidak B. L., Olsen G. J., Larsen N., Overbeek R., McCaughey M. J.,
Woese C. R. 1996. The ribosomal database project (RDP). Nucleic
Acids Res. 24:82–85.
Meacham C. A. 1994. Phylogenetic relationships at the basal radiation
of angiosperms: further study by probability of character compati-
bility. Syst. Bot. 19:506–522.
Mooers A., Holmes E. 2000. The evolution of base compo-
sition andphylogeneticinference.
365–369.
Olsen G. 1987. Earliest phylogenetic branchings: comparing rRNA-
based evolutionary trees inferred with various techniques. Cold
Spring Harbor. Symp. Quant. Biol. 52:825–837.
TrendsEcol.Evol. 15:
Olsen G. J., Matsuda H., Hagstrom R., Overbeek R. 1994. fastDNAmL:
a tool for construction of phylogenetic trees of DNA sequences us-
ing maximum likelihood. Comput. Appl. Biosci. 10:41–48.
Olsen G., Pracht S., Overbeek R. 1998. DNArates. Version 1.1.
Philippe H., Zhou Y., Brinkmann H., Rodrigue N., Delsuc F. 2005. Het-
erotachy and long-branch attraction in phylogenetics. BMC Evol.
Biol. 5:50.
Pisani D. 2004. Identifying and removing fast-evolving sites using
compatibility analysis: an example from the Arthropoda. Syst. Biol.
53:978–989.
Puigbo P., Garcia-Vallve S., McInerney J. O. 2007. TOPD/FMTS: a
new software to compare phylogenetic trees. Bioinformatics. 23:
1556–1558.
Rambaut A., Grassly N. C. 1997. Seq-Gen: an application for the Monte
Carlo simulation of DNA sequence evolution along phylogenetic
trees. Comput. Appl. Biosci. 13:235–238.
Ruvolo M. 1997. Molecular phylogeny of the hominoids: inferences
from multiple independent DNA sequence data sets. Mol. Biol.
Evol. 14:248–265.
Satta Y., Klein J., Takahata N. 2000. DNA archives and out nearest
relative: the trichotomy problem revisited. Mol. Phylogenet. Evol.
14:259–275.
Schmidt H., Strimmer K., Vingron M., von Haeseler A. 2002. TREE-
PUZZLE: maximum likelihood phylogenetic analysis using quar-
tets and parallel computing. Bioinformatics. 18:502–504.
Schwartz J. 1984. The evolutionary relationships of man and orang-
utans. Nature. 308:501–505.
Shoshani J., Groves C., Simons E., Gunnell G. 1996. Primate phy-
logeny: morphological vs molecular results. Mol. Phylogenet. Evol.
5:102–154.
Strimmer K., von Haeseler A. 1997. Likelihood-mapping: a simple
method to visualize phylogenetic content of a sequence alignment.
Proc. Natl. Acad. Sci. U.S.A. 94:6815–6819.
Townsend J. P. 2007. Profiling phylogenetic informativeness. Syst. Biol.
56:222–231.
Wilgenbusch J. C., Swofford D. 2003. Inferring evolutionary trees with
PAUP*. Curr. Protoc. Bioinformatics. Chapter 6: Unit 6.4.
Wilkinson M. 1998. Split support and split conflict randomization tests
in phylogenetic inference. Syst. Biol. 47:673.
Yang Z. 1993. Maximum-likelihood estimation of phylogeny from
DNA sequences when substitution rates differ over sites. Mol. Biol.
Evol. 10:1396–1401.
Yang Z. 1996. Among-site rate variation and its impact in phylogenetic
analyses. Trends Ecol. Evol. 11:367–372.
at National University of Ireland, Maynooth on October 17, 2011
sysbio.oxfordjournals.org
Downloaded from