Journal of Heredity 2011:102(1):130–138 Ó The American Genetic Association. 2010. All rights reserved.

doi:10.1093/jhered/esq092 For permissions, please email: [email protected].

Advance Access publication August 8, 2010

Recent Trends in Molecular

Phylogenetic Analysis: Where to Next?
CHRISTOPHER BLAIR AND ROBERT W. MURPHY
From the Department of Ecology and Evolutionary Biology, University of Toronto, Toronto, ON M5S 3B2, Canada
(Blair and Murphy); and the Department of Natural History, Royal Ontario Museum, Toronto, ON M5S 2C6, Canada
(Blair and Murphy).

Downloaded from https://academic.oup.com/jhered/article/102/1/130/821529 by guest on 27 September 2022

Address correspondence to C. Blair at the address above, or e-mail: [email protected].

The acquisition of large multilocus sequence data is methods are showing exceptional promise for these types of
providing researchers with an unprecedented amount of data but have not yet garnered sufficient attention among
information to resolve difficult phylogenetic problems. With empiricists. As phylogenetic methods change rapidly and are
these large quantities of data comes the increasing challenge often complex, we hope this review can serve as a guideline
regarding the best methods of analysis. We review the for researchers new to the field. This review will also be of
current trends in molecular phylogenetic analysis, focusing use to current phylogeneticists looking for novel methods to
specifically on the topics of multiple sequence alignment and analyze their multilocus data sets. We stress that this review
methods of tree reconstruction. We suggest that traditional is by no means exhaustive. However, we feel that the topics
methods are inadequate for these highly heterogeneous discussed are likely those that researchers will encounter in
data sets and that researchers employ newer more modern molecular systematics.
sophisticated search algorithms in their analyses. If we are
to best extract the information present in these data sets,
a sound understanding of basic phylogenetic principles Alignment
combined with modern methodological techniques are
A robust phylogenetic hypothesis is intricately linked to the
necessary.
quality of the MSA. The accurate determination of character
Key words: alignment, Bayesian, mixture modeling, parsimony, homology is vital to infer evolutionary history, and thus,
species TREE substantial effort has been placed on finding optimal
methods to align divergent orthologous sequences. We
define homology here as identity in character state due to
With third-generation sequencing technology rapidly ap- shared evolutionary history. Due to intronic and intergenic
proaching, it will become more feasible to obtain large regions, accurate assessments of base-pair homology will
multilocus data sets to infer evolutionary relationships become even more problematic as larger multilocus data
(Genome 10k Community of Scientists 2009). These sets become available. Traditionally, an MSA is performed
enormous quantities of data have spawned the development either manually or with the aid of different computer
of several new programs for phylogenetic inference for algorithms. Manual alignments (i.e., by eye) are often
these highly heterogeneous data sets. From multiple relatively straightforward when inferring relationships using
sequence alignment (MSA) to species tree construction, highly conserved coding sequences. However, performing
these new methods are changing the way we gather and manual alignment of noncoding sequences can be prob-
manipulate data and analyze and interpret results. The lematic as insertion/deletion (indel) events are more
purpose of this paper is to provide a brief review of both common and do not necessarily occur in multiples of
traditional and more modern methods available to analyze 3 bases.
diverse data sets in the era of multilocus phylogenetics. We To try to reduce the degree of subjectivity introduced
start by providing a discussion of methods to infer an into an MSA of distantly related groups, many researchers
alignment of orthologous sequences and ways to extract turn to computer software packages. Traditionally, the most
information present in insertion/deletion events. This is common programs used to construct an MSA were the
followed by a review of the often-controversial direct Clustal series (Thompson et al. 1994; Larkin et al. 2007).
optimization (DO) techniques. The last 2 sections specif- Clustal is a progressive method that uses a neighbor joining
ically review methods of tree reconstruction, the first of (NJ) guide tree to create the MSA from a series of pairwise
which focuses mostly on concatenation; a method com- alignments. The user can specify specific parameters before
monly used in multilocus studies. In contrast, species tree the analysis is run, including the costs of both inserting a gap

130
 Perspectives

and extending a previously inserted gap. Although Clustal missing data (most cases) or coded as a fifth character state
can be extremely helpful when working with difficult (Swofford et al. 1996). Both of these methods are potentially
sequences, the resulting MSA is often edited manually to problematic in that the former completely discards relevant
subjectively improve character homology. However, oppo- evolutionary information (Simmons et al. 2001; Kawakita
nents of manual alignment methods argue that introducing et al. 2003), whereas the latter assumes that gaps represent
researcher bias into an analysis yields the resulting alignment independent evolutionary events; a highly unlikely scenario
and subsequent phylogenetic hypothesis nonreproducible. (Simmons and Ochoterena 2000). These issues also
The acquisition of large heterogeneous data sets will extended into probabilistic phylogenetic inference in that
result in orthologous sequences containing several distinct parameters were estimated without taking indel events into
indel events in noncoding regions. Thus, a refinement of account.
methods used to create an MSA is necessary if we are to To adequately extract the phylogenetic information
accurately determine base-pair homology (Wallace et al. present in indels (gaps), researchers are beginning to look
2005). Over the past decade, several studies have quantified at both alternative means of character coding (Simmons and

Downloaded from https://academic.oup.com/jhered/article/102/1/130/821529 by guest on 27 September 2022

the accuracy of common MSA methods using simulated Ochoterena 2000; Simmons et al. 2007) as well as the
data sets (Lassmann and Sonnhammer 2002; Katoh et al. formulation of indel models of sequence evolution (Thorne
2005; Edgar and Batzoglou 2006; Nuin et al. 2006). For et al. 1991, 1992; Miklós et al. 2004). Simmons and
example, a recent study compared the performances of 6 Ochoterena (2000) suggested using both simple and
alignment packages to infer the correct MSA from complex coding of indels to quantify synapomorphies
sequences containing gaps (Golubchik et al. 2007). Gaps present between taxa. Simple indel coding assumes that
are inserted into an MSA to account for indel events, but the each contiguous stretch of gaps with identical 5# and 3#
location of gap insertion is often ambiguous at best. They termini comprises a single evolutionary event. A binary
found that the more commonly used progressive packages (presence/absence) character state matrix is then concate-
like Clustal and T-Coffee (Notredame et al. 2000) nated to the remaining alignment and analyzed together
performed quite poorly when indels were present. under any optimality criterion. For probabilistic inference,
Conversely, programs such as DIALIGN-T (Morgenstern the morphological models presented in Lewis (2001) can be
2004) and MAFFT (Katoh et al. 2002) often recovered used for the binary matrix as a separate partition. Indel
a more accurate MSA. However, combining results from coding can be performed manually or automated using
different alignment algorithms using M-Coffee (Wallace programs such as GapCoder (Young and Healy 2003) and
et al. 2006) and/or removing ambiguously aligned regions BARCOD (Barriel 1994). Studies have shown that including
with Gblocks (Talavera and Castresana 2007) may result in the information present in gaps can: 1) dramatically increase
more robust phylogenetic hypotheses, especially when the number of potentially phylogenetically informative
alignments differ substantially with different algorithms. characters and 2) lead to a more strongly supported
Probabilistic methods of MSA are also beginning to phylogenetic hypothesis (Giribet and Wheeler 1999;
determine character homology with noncoding DNA. These Kawakita et al. 2003; Ogden and Rosenberg 2007b).
methods rely on statistical models of sequence evolution to Furthermore, indel characters tend to show reduced levels
obtain the alignment or set of alignments with the highest of homoplasy as compared with nucleotide characters
probability given a set of orthologous sequences. Early (Simmons et al. 2001). Thus, researchers should continue
models were only able to accommodate indels one base pair to seek to maximize the evolutionary information present in
in length and are thus unrealistic for most molecular data indel events, especially as more noncoding sequences are
sets (Thorne et al. 1991, 1992). However, more realistic used and more realistic models of evolution are developed.
indel models continue to be developed specifically for
noncoding sequences (Keightley and Johnson 2004). The
Direct Optimization
model used in the package MCALIGN (Keightley and
Johnson 2004), for example, is parameterized by comparing An alternative to constructing an MSA prior to phylogenetic
the empirical distribution of indel lengths and frequencies inference is to use DO procedures. DO is different from
relative to the rate of nucleotide substitution. Although the other approaches in that the alignment and phylogenetic
method is potentially promising, its current implementation tree are estimated simultaneously. Optimization can be
only allows for the alignment of 2 or 3 orthologous performed either under a parsimony or under a probabilistic
sequences. framework. The program POY (Wheeler 1996; Varón et al.
2009), for example, estimates both the phylogenetic tree and
the best alignment based on the MP criterion. Previous
Gap Coding
versions of POY were also able to implement DO in
Following the construction of an MSA for the traditional a likelihood framework. Newer programs such as StatAlign
2-step MSA þ phylogeny estimation procedure, the (Novák et al. 2008), BAli-Phy (Suchard and Redelings 2006),
researcher is left with the decision of how to handle the and BEAST (Lunter et al. 2005) incorporate models of
gaps inserted into the data set by the MSA algorithm to sequence evolution to estimate the posterior distribution of
account for indel events. For most traditional maximum a set of trees and alignments based on Bayesian inference
parsimony (MP) analyses, gaps have been either coded as (BI). The Bali-Phy software shows exceptional promise in

131
Journal of Heredity 2011:102(1)

that its models allow for nested or overlapping indel events data exploration and analysis and are currently used less
(Redelings and Suchard 2005, 2007), whereas other methods frequently in phylogenetic analysis than more utilitarian
utilize the more common TKF1 and TKF2 indel models software packages.
(Thorne et al. 1991, 1992). However, joint estimation of Historically, the most widely used program to estimate
alignment and phylogeny in a probabilistic framework is trees under the parsimony criterion is PAUP* or phyloge-
currently computationally intensive and feasible only with netic analysis using parsimony (*and other methods)
smaller data sets (Lunter et al. 2005). These methods also fit (Swofford 2002). This program has several advantages to
a single model to the data, which may not be justified with the user, some of which include the ability to analyze
multilocus data sets. molecular or morphological data or data sets containing
Although DO procedures can provide an alternative to a mixture of character types. PAUP* is also able to analyze
constructing a subjective MSA, there are other technical and data under different criteria including the NJ and maximum
philosophical factors that possibly confound their utility likelihood (ML) methods. Compared with several other
(Simmons 2004; Simmons and Ochoterena 2000). Studies phylogenetic software packages, the Macintosh version of

Downloaded from https://academic.oup.com/jhered/article/102/1/130/821529 by guest on 27 September 2022

have compared the performance of DO with POY versus PAUP* is also quite user-friendly, which has added to
traditional Clustal MSA þ MP in PAUP* (Swofford 2002) researchers preference for this program versus other
and found the former to be inferior in most of the analyses parsimony software. Other functions in PAUP* that have
performed (Kjer et al. 2007; Ogden and Rosenberg 2007a). made it a versatile choice for phylogenetic inference are: 1)
However, a more recent simulation study showed that these the ability to implement a variety of different models of
results are due, in part, to the use of simple gap penalties in evolution in a distance or likelihood framework; 2) the
POY (Liu, Nelesen, et al. 2009). Conversely, using an calculation of heterogeneous patterns in data with the
affine-gap penalty in POY often leads to more accurate partition homogeneity test or incongruence length differ-
alignment and phylogenetic hypotheses than the tradi- ence test (Farris et al. 1994); 3) the detection of saturation;
tional Clustal þ MP approach. By affine penalty, we refer to and 4) the calculation of pairwise base differences with or
the additive nonzero cost of both inserting a gap and without accounting for multiple hits.
extending a previously inserted gap. For example, under Although the benefits of PAUP* are numerous, the
a simple gap penalty, the cost incurred would equal kC, program suffers from severe limitations when data sets get
where C equals the gap opening cost and k equals the too large, especially when there is a lack of strong
length of the gap. Under an affine penalty, a gap of length phylogenetic signal. Furthermore, it is well known that
k would cost Copen þ kCextend, where Copen equals the cost for implicit enumeration and branch-and-bound meth-
of inserting a gap and kCextend is the cost of extending the ods, having a data set with more than 20 taxa can take an
inserted gap of length k. Few comprehensive simulation enormous amount of time and computational power that is
studies have investigated the performance of likelihood- generally not feasible with modern computer technology
based DO methods, but it is anticipated that the power of (Swofford et al. 1996). Instead, performing a heuristic
these methods will continue to increase as more realistic search, usually with a set of random addition sequences
models of evolution become incorporated into the analysis (RASs) followed by tree bisection–reconnection (TBR)
(Miklós et al. 2004). branch swapping is implemented in hopes of finding a set
of most parsimonious trees. However, even these heuristic
methods can take an inordinate amount of time when data
Phylogenetic Inference sets get too large (both in taxa and base pairs). This makes
PAUP* an unlikely candidate to analyzing larger multilocus
Concatenation
data sets in a parsimony framework. Furthermore, providing
Any method of data analysis must rely on an underlying set support for nodes via bootstrapping (Felsenstein 1985) or
of assumptions. A primary objective of any scientific jackknifing (Lanyon 1985) would be a computational burden
investigation is to minimize the number of unnecessary with these quantities of data.
and unjustifiable assumptions required to explain patterns in A relatively new phylogenetic inference package, TNT
the data. In this regard, MP analysis continues to be (Tree Analysis Using New Technology; Goloboff et al.
a popular cladistic method to reconstruct evolutionary 2003, 2008) is showing exceptional promise for the analysis
histories. Although not statistical in the sense that the of large data sets (.150 taxa) under MP (Hovenkamp 2004;
criterion does not impose a probabilistic model of sequence Giribet 2005; Meier and Ali 2005; Goloboff and Pol 2007).
evolution, many cladists believe that MP outperforms TNT is available for Linux, Macintosh, and Windows
model-based methods of tree reconstruction (including platforms (GUI version for Windows only) with a down-
distance-based phenetic methods). This is due, in part, to loadable PowerPoint presentation that explains the basic
factors such as the additional assumptions involved with features of the program. TNT’s implementation of 4 ‘‘New
phenetic and likelihood-based analyses (Siddall and Kluge Technology’’ search algorithms (sectorial searches, tree
1997; Siddall 2001). Several computer programs are able to drifting, tree fusing, and the ratchet) provides researchers
perform MP analysis using morphological and/or molecular with novel and faster methods to search for optimal trees.
data (Hennig86, Farris 1988; NONA/Pee-Wee, Goloboff TNT allows for the analysis of data sets that would normally
1994) Many of these programs are limited in their scope of require between 300 and 900 times longer with traditional

132
 Perspectives

phylogenetic programs such as PAUP*(Goloboff 1999). For researchers as to the best way to determine what the optimal
a comprehensive explanation of these search algorithms, the partitioning strategy is for a particular data set. This problem
reader is directed to the original papers (Goloboff 1999; is further exacerbated by the fact that overpartitioning a data
Nixon 1999). In addition to these new algorithms providing set can be just as harmful as underpartitioning and can
a more thorough evaluation of tree space with larger data introduce several unnecessary additional parameters to the
sets, simulation studies have shown that standard RAS þ analysis that may or may not be justified by the data (Brown
TBR searches in TNT take a fraction of time as those and Lemmon 2007). The use of Bayes factors (Kass and
implemented in PAUP*. Indeed it is shown that TNT is able Raftery 1995) is currently the most common way to assess
to perform rapid parsimony analysis with a high number of the optimal partition strategy for a given data set. However,
taxa (;1000; Goloboff 1999), although it is not yet known there is still some degree of subjectivity in the calculation
how the algorithms would handle data sets composed of and interpretation of values. For example, some researchers
thousands to tens of thousands of base pairs. More advocate a significance value of 10 or greater (Brandley et al.
simulation studies are needed to determine how TNT will 2005), whereas others suggest that any value above 2 is

Downloaded from https://academic.oup.com/jhered/article/102/1/130/821529 by guest on 27 September 2022

handle these data sets for a greater number of taxa. significant (Raftery 1996). This means that when comparing
However, it is highly likely that TNT will serve as a prime 2 separate models, model one versus model 2 (BF12),
candidate to quickly and efficiently analyze multilocus data a 2(lnBF) . 10 favors the former over the latter. The values
under the parsimony criterion. used for the calculation of BF are typically the marginal
Phylogenetic packages using probabilistic methods such likelihoods of each run, which can be calculated in software
as ML and BI have gained popularity due to their ability to packages such as Mathematica or approximated by the
incorporate various models of sequence evolution to best harmonic mean values in MrBayes.
explain the data. Historically, ML methods were used as As sequencing technology progresses, heterogeneous
a parametric comparison with MP analysis. Over the past data sets encompassing various genomic regions including
2 decades, its use has diminished as newer Bayesian protein-coding genes, rRNA genes, intergenic regions, and
techniques have become available that are better able to introns will become standard practice. With these diverse
exploit the information contained in a data set. For example, data sets, it will be necessary to determine the best methods
early ML methods (i.e., PAUP*) were more computationally to account for both the pattern as well as rate variation in
demanding than Bayesian methods, especially when applying nucleotide substitution present throughout the different
more than one model of evolution to the data. Indeed, most regions. Although partitioning the data in MrBayes gener-
researchers who used the ML criterion to generate ally leads to improved likelihood scores, this usually comes
a phylogenetic hypothesis fit a single model of evolution to at the cost of increased computation time required to
the data, even if the data were composed of many different reach convergence (Roberts et al. 2009). Consider a data
genes and codons that are likely evolving at very different set composed of 3 unlinked protein-coding genes. As-
rates. Recently, newer, more sophisticated ML algorithms signing one model to the entire data set may be relatively
such as those implemented in GARLI (Zwickl 2006) and computation-friendly, but it may not be philosophically
RAxML (Stamatakis 2006) have been developed for rapid justifiable based on the variance in evolutionary rate and
likelihood analysis. RAxML in particular is showing pattern. It may be more suitable to divide the data into
exceptional promise for phylogenetic inference with thou- discrete partitions (gene and codon within gene) to fully
sands of taxa and/or base pairs. RAxML is also able to accommodate the heterogeneous rates and patterns most
accommodate mixed model ML analysis (GARLI is currently likely present in the data. With 3 protein-coding genes,
unable to do so), fitting independent evolutionary models to however, this would require 9 separate partitions, with
alignments of both nucleotides and amino acids. With these parameters estimated separately for each partition (assuming
improved algorithms, ML is again becoming a popular choice an unlinked model). Different priors on branch lengths and
for phylogenetic analysis of exceptionally large data sets. evolutionary rates can also be specified for each data
Bayesian methods of phylogenetic inference continue to partition before the analysis is implemented (Brown et al.
be among the most popular choice for analyzing large 2010). However, increasing the number of partitions and
complex data sets composed of different genic regions. The parameters increases computation time and parameter
majority of past and current studies implement a partitioned variance at an exponential rate (Roberts et al. 2009).
Bayesian analysis in the program MrBayes (Ronquist and Additional partitions also decrease the number of sites
Huelsenbeck 2003) dividing the data a priori based on (characters) per partition, increasing the systematic error in
different functional regions such as genes, codon position in model fitting (Brandley et al. 2005; Roberts et al. 2009).
protein-coding genes, and stems and loops in ribosomal Highly partitioned data sets may also lead to errors in
RNA (rRNA) (Nylander et al. 2004; Brandley et al. 2005). branch length estimation (Marshall et al. 2006) and influence
Simulation and empirical data sets have shown that convergence and mixing of the Markov chains, which may
partitioning the data into discrete regions and assigning require changing the heating scheme (Beiko et al. 2006).
different evolutionary models to each region significantly Although certainly feasible with a small to moderate number
improves the marginal likelihood values for the posterior of loci, applying these methods to large data sets may lead to
distribution (Nylander et al. 2004; Brown and Lemmon erroneous conclusions due to inappropriate models and
2007). However, there has been less consensus among prior distributions.

133
Journal of Heredity 2011:102(1)

A newer and more promising approach to BI of alignment at a specific location in the tree (Fitch and
multilocus data sets can be found in the software Markowitz 1970; Pagel and Meade 2008). Heterotachy is
package BayesPhylogenies (Pagel and Meade 2004). The a common phenomenon in molecular evolution but is
program implements a mixture model to account for pattern seldom incorporated into a molecular phylogenetic analysis
heterogeneity present in diverse data sets and assigns (Lopez et al. 2002; Philippe et al. 2005). BayesPhylogenies
a model to each site in an alignment of nucleotide accomplishes this by assigning more than one branch length
sequences. BayesPhylogenies uses the likelihood function to a specific region in a phylogenetic tree. These additional
to calculate how many independent substitution rate branch length sets can be specified by the user a priori or
matrices, Q, are needed to best explain the data without estimated directly from the data by the RJ procedure (Pagel
the need for a priori partitioning of the data. Each rate and Meade 2008). Using the RJ procedure to estimate the
matrix is composed of a known model of evolution, likelihood of additional branch lengths has numerous
generally corresponding to the general time reversible benefits over assigning additional complete sets to the
(GTR) model with or without gamma rate variation (Yang entire data a priori (Pagel and Meade 2008). For example,

Downloaded from https://academic.oup.com/jhered/article/102/1/130/821529 by guest on 27 September 2022

1994), and parameters are estimated separately for each assigning additional sets forces more than one branch
matrix. The GTR method is usually used for each matrix length for every branch of a tree, most of which are
because it allows for the most free parameters. In a unjustified and add unnecessary parameters to the analysis.
traditional likelihood framework, we seek to maximize the The RJ procedure, on the other hand, only assigns addi-
likelihood of observing the data given a specific set of tional branch lengths where the data suggest that they are
parameters. This can be written as P(D|Q, T), where D is necessary, in turn minimizing the number of additional
an alignment of nucleotide sequences, Q corresponds to parameters required to best explain the data.
a model of sequence evolution, and T is a phylogenetic tree. The benefits of employing a mixture model Bayesian
The mixture model method is able to incorporate additional analysis over a partitioned analysis are relatively straightfor-
models to maximize the likelihood and can be written as ward. First, these models do not require a priori partitioning
P(D|Q1, Q2, Q3, . . . ,Qj, T), where Q1–Qj represents how of the data, which is often a subjective matter. Second,
many models (rate matrices) are required to best explain the mixture model methods can select the best modeling
data. Simulation and empirical analyses have shown strategy a posteriori based on the given data (RJ method),
that mixture models can outperform mixed model BI eliminating the need for model selection criteria such as
(i.e., partitioning) based on increased likelihood scores, Bayes factors, AIC, and Bayesian Information Criteria.
usually without the incorporation of additional unnecessary Third, mixture models allow for within partition pattern
parameters (Pagel and Meade 2004, 2005). variability, which is quite common in an alignment of
Although employing a mixture model analysis may be molecular sequences but seldom incorporated (Pagel and
more suitable to a particular data set, the researcher often Meade 2004). Finally, mixture models have the ability to
does not know a priori how many independent rate reduce node-density artifacts (i.e., underestimate of the
matrices are required to best explain the data. One method amount of change in long branches compared with shorter
that can be used to determine the minimum number of rate branches with more nodes) often present in a data set
matrices is to simply repeat the analysis several times (Venditti et al. 2008). For these reasons, as well as those
specifying a different number of Q matrices with and presented above, researchers are beginning to incorporate
without gamma rate variation (C ) and separate base new heterogeneous models in empirical data sets composed
frequencies (p) and use model selection criteria such as of several independent loci. Over the past few years, mixture
Bayes factors or Akaike’s information criterion (AIC) to models have been used to infer phylogeny in several groups
select the best strategy (Roberts et al. 2009). The researcher including mammals (Roberts et al. 2009), flies (Lewis et al.
may also plot the likelihood returns for each separate 2005), and lizards (Zaldivar-Riverón et al. 2008; Schulte and
run versus Q (with and without C ) in order to visually Cartwright 2009). These studies find mixture model results
assess stationarity (see e.g., Pagel and Meade 2004). comparable with traditional probabilistic inference methods,
Alternatively, the reversible jump (RJ) procedure (Green although likelihood scores are generally higher. However,
1995; Huelsenbeck et al. 2004) can accommodate un- methods to assess adequate mixing and convergence for
certainty by allowing the joint estimation of parameters in mixture models are still in development. Additional
the posterior distribution. Along with estimating the empirical studies will allow us to draw better conclusions
parameters of interest such as tree topology and branch as to the congruence of mixture model approaches with
lengths, the RJ method also estimates the minimum number more traditional partition-based methods.
of rate matrices required to maximize the likelihood of It is predicted that mixture models will begin to
observing the data. This method is advantageous as it does be incorporated more frequently as multilocus data sets
not to make any a priori assumptions regarding the pattern become available. This is due to a combination of com-
of sequence evolution. putational efficiency and philosophical justifications to
Although computationally intensive, an additional useful handle highly heterogeneous data sets. It becomes prob-
benefit of employing a mixture model analysis is that the lematic to fit an inordinate number of evolutionary models
program is able to incorporate heterotachy; variation in the to a data set containing tens to hundreds of independent loci
rate of evolution through time in a specific region of an a priori and running a partitioned analysis. These models will

134
 Perspectives

not only add an enormous amount of additional parameters total evidence topology. These include partitioned Bremer
to the analysis but may also prove to be a poor fit to the data support (Baker and DeSalle 1997), partitioned likelihood
due to overparameterization. Estimating the number of support (Lee and Hugall 2003), and the partition addition
required models directly from the data is a much more bootstrap alteration (Struck et al. 2006). Given the total
feasible and defensible goal. Consider a data set containing evidence topology, these methods quantify the support or
100 000 bp spanning several protein coding exons, introns, conflict for each node calculated from each partition in either
and intergenic regions (soon to be a moderately sized data a parsimony or a likelihood framework. Indeed, it is the
set). If the user wished to perform a partitioned Bayesian concatenation approach that has garnered the most attention
analysis, these data would have to be divided a priori into an from systematists over the past few decades and has proved
unmanageable number of partitions, increasing both the effective in some genomic data sets (e.g., Rokas et al. 2003).
number of assumptions and computational burden. Con- Although the concatenation method may provide more
versely, using the RJ mixture model procedure to estimate phylogenetic resolution simply by increasing the number
the number of required rate matrices directly from the data of informative characters, information about variance in

Downloaded from https://academic.oup.com/jhered/article/102/1/130/821529 by guest on 27 September 2022

may show that the data can be explained by only a few gene coalescence is lost. This has spawned an exciting new
different rate matrices. This in turn drastically reduces the era of research into methods to 1) account for gene tree
number of required parameters in the analysis and will discordance in estimating a species tree and 2) move away
provide higher likelihood scores. from concatenation (Edwards et al. 2007; Kubatko and
Degnan 2007; Knowles 2009).
It is currently assumed that incomplete lineage sorting
Species Trees
due to deep coalescence is the most common mechanism
The reconciliation of a well-supported species tree is the leading to incongruence among gene trees (Degnan and
primary interest in systematics. Traditionally, the steps Rosenberg 2009). Thus, most of the recent species tree
involved in a molecular phylogenetic investigation included methods do not account for the other potential factors
1) sequencing a single gene from a representative individual outlined above (but see Kubatko 2009). Generally speaking,
for each taxon under study; 2) constructing a gene tree of species tree methods are classified as either summary
the sequences using available phylogenetic methods; and 3) statistics based or probabilistic based (Knowles 2009;
equating the resolved gene tree with the species tree. Liu, Yu, et al. 2009; McCormack et al. 2009). Early summary
Although these gene trees often serve as an adequate statistics approaches such as minimizing deep coalescence
hypothesis for the species tree in older groups, discor- (Maddison and Knowles 2006) are computationally trac-
dance between gene trees can be quite common, especially table but often less robust than probabilistic methods
for rapid speciation events and recently diverged taxa (McCormack et al. 2009). However, newer summary
(Degnan and Rosenberg 2006, 2009; Carstens and Knowles statistics methods that account for coalescence variance
2007). As we are moving away from single-locus phyloge- such as STAR and STEAC (Liu, Yu, et al. 2009) are showing
netics to more multilocus studies (Edwards 2008; Brito exceptional promise to construct species trees. These
and Edwards 2009), gene tree heterogeneity will become methods are relatively quick and can be implemented in
a common phenomenon in molecular investigations. the statistical package R.
Extracting the phylogenetic signal from these discordant ML and Bayesian methods have been developed to
gene trees is paramount in hypothesizing the most likely model gene coalescence to infer the most likely species tree.
species tree. Therefore, several recent methods have been These heavily parameterized models allow for the stochas-
proposed to estimate the most likely species tree from a set ticity of coalescence times between gene lineages to be
of independent gene trees (reviewed in Knowles 2009). integrated into species tree inference probabilistically. In
It is now well understood that gene tree heterogeneity is a fully Bayesian framework, for example, the programs
ubiquitous (Degnan and Rosenberg 2009). Factors such as BEST (Bayesian Estimation of Species Trees; Liu 2008; Liu
horizontal gene transfer (Doolittle 1999), gene duplication and Pearl 2007) and *BEAST (Heled and Drummond 2010)
(Page and Charleston 1997), recombination, hybridization use multiple alleles per species to estimate the posterior
and introgression (Rieseberg et al. 2000), and incomplete distribution of gene and species trees. Prior distributions of
lineage sorting (Pamilo and Nei 1988) can result in gene trees gene trees are conditioned on the species tree in which the
that are discordant with other gene trees as well as the species gene trees are embedded. As a by-product of the joint
tree (Maddison 1997). Several traditional and more recent estimation of gene trees and the species tree, the programs
methods have been used to deal with this discordance in also provide divergence times and ancestral population sizes.
multilocus data sets. For example, the most common way to The latter has been shown to rely heavily on the prior
incorporate information from multiple genes is to concat- distribution of h 5 4Nel, where Ne is the effective
enate them into one supermatrix and analyze them as one population size and l is the per-site mutation rate (Liu and
under the total evidence approach (Kluge 1989). Concate- Pearl 2007). Although species tree inference with BEST and
nation assumes that the phylogenetic signal present in most *BEAST have been shown to outperform concatenation,
partitions will swamp out the conflicting signal in others. computational demand currently limits their utility when
Over the years, several methods have been developed to applying the methods to larger data sets (Cranston et al.
examine the contribution of individual gene partitions to the 2009; Liu, Yu, et al. 2009).

135
Journal of Heredity 2011:102(1)

Because joint estimation of gene and species trees is drawing from expertise across diverse backgrounds. With
often intractable with large quantities of data, other ML and new sequencing technologies upon us, and new methods of
Bayesian methods have been employed. Bayesian concor- analysis, simulation studies are vital if we are to ascertain the
dance analysis (BCA; Ané et al. 2007), for example, uses best approach to phylogenetic inference. This simulation-
individually calculated gene trees to infer the species tree based research has thus been a necessary and rapidly
that maximizes the bipartition concordance among each progressing component of multilocus phylogenetics and
gene tree. BCA can be implemented in the program BUCKy much more work is needed. Furthermore, little work has
(Ané et al. 2007) and requires a set of calculated gene trees been done to quantify the application of new methods to
along with a single user-defined input parameter a that phylogeographic studies. We are approaching an exciting
quantifies the degree of gene tree incongruence. BCA has new era of molecular systematics and as more heteroge-
been shown to be a highly accurate alternative to the neous data sets begin to accumulate, the efficacy of current
demanding algorithm used in BEST and the method is methodology is sure to be put to the test.
generally insensitive to values of a (Cranston et al. 2009).

Downloaded from https://academic.oup.com/jhered/article/102/1/130/821529 by guest on 27 September 2022

A promising ML method is found in the program STEM
(Species Tree Estimation Using Maximum Likelihood; Funding
Kubatko et al. 2009). Like BCA, STEM uses the calculated
Natural Sciences and Engineering Research Council of
gene trees (assuming a molecular clock) as input data to
Canada Discovery Grant A3148 to R.W.M.
calculate the most likely species tree under coalescence. The
program can also evaluate the likelihood of a user-specified
input tree as well as search for a set of high likelihood
topologies. Prior to analysis, the user must specify a value
Acknowledgments
for h (to convert branch lengths into coalescent units) and We thank the graduate students and faculty of the Royal Ontario Museum
a per-locus evolutionary rate r. Although ML inference with for their discussions and insightful criticisms. We thank Ida M. Conflitti and
3 anonymous reviewers for their comments to improve the manuscript.
STEM has been shown to improve over concatenation ML
analysis (Kubatko et al. 2009), it is still unknown how robust
the method is to starting parameter values.
References
Ané C, Larget B, Baum DA, Smith SD, Rokas A. 2007. Bayesian estimation
of concordance among gene trees. Mol Biol Evol. 24:412–426.
Conclusions Baker RH, DeSalle R. 1997. Multiple sources of character information and
We hope that this review encourages both empirical and the phylogeny of Hawaiian drosophilids. Syst Biol. 46:654–673.
theoretical systematists to strongly consider the status and Barriel V. 1994. Molecular phylogenies and nucleotide insertion-deletion.
future of phylogenetic inference methods. As multilocus CR Acad Sci III. 317:693–701.
data sets become the norm across laboratories, some of the Beiko RG, Keith JM, Harlow TJ, Ragan MA. 2006. Searching for convergence
most commonly employed techniques for both MSA and in phylogenetic Markov chain Monte Carlo. Syst Biol. 55:553–565.
tree reconstruction will no longer be adequate for generating Brandley MC, Schmitz A, Reeder TW. 2005. Partitioned Bayesian analyses,
phylogenetic hypotheses. Instead, alternate and more partition choice, and the phylogenetic relationships of scincid lizards.
sophisticated search algorithms are required in order to Syst Biol. 54:373–390.
fully exploit the information contained in these large Brito P, Edwards SV. 2009. Multilocus phylogeography and phylogenetics
quantities of data. As highly heterogeneous data sets using sequence-based markers. Genetica. 135:439–455.
become available, testing the accuracy of both modern Brown JM, Hedtke SM, Lemmon AR, Lemmon EM. 2010. When trees
alignment algorithms and DO methods through simulation grow too long: investigating the causes of highly inaccurate Bayesian
branch-length estimates. Syst Biol. 59:145–161.
will become even more important.
For traditional phylogenetic inference, MP analysis will no Brown JM, Lemmon AR. 2007. The importance of data partitioning and the
utility of Bayes factors in Bayesian phylogenetics. Syst Biol. 56:643–655.
doubt continue to play a role. In this regard, TNT is showing
promise for dealing with difficult phylogenetic problems. Brudno M, Do CB, Cooper GM, Kim MF, Davydov E, NISC Comparative
Sequencing Program, Green ED, Sidow A, Batzoglou S. 2003. LAGAN and
Furthermore, model-based concatenation methods using Multi-LAGAN: efficient tools for large-scale multiple alignment of genomic
mixture models in BayesPhylogenies seem promising for DNA. Genome Res. 13:721–731.
multilocus data sets. However, there have been few Carstens BC, Knowles LL. 2007. Estimating species phylogeny from gene-
simulations to quantify the accuracy of the model compared tree probabilities despite incomplete lineage sorting: an example from
with other methods including direct species tree inference. Melanoplus grasshoppers. Syst Biol. 56:400–411.
Several methods outlined in this review are also showing Cranston KA, Hurwitz B, Ware D, Stein L, Wing RA. 2009. Species trees
promise for analyzing phylogenomic data sets, including the from highly incongruent gene trees in rice. Syst Biol. 58:489–500.
alignment algorithms in the programs Multi-LAGAN Darling ACE, Mau B, Blatter FR, Perna NT. 2004. Mauve: multiple
(Brudno et al. 2003) and MAUVE (Darling et al. 2004) alignment of conserved genomic sequence with rearrangements. Genome
and the species tree methods found in programs such as Res. 14:1394–1403.
BCA, STEM, and STEAC. The field of molecular Degnan JH, Rosenberg NA. 2006. Discordance of species trees with their
systematics has always been a rapidly evolving discipline most likely gene trees. PloS Genet. 2:762–768.

136
 Perspectives

Degnan JH, Rosenberg NA. 2009. Gene tree discordance, phylogenetic Kjer KM, Gillespie JJ, Ober KA. 2007. Opinions on multiple sequence
inference, and the multispecies coalescent. Trends Ecol Evol. 24:332–340. alignment, and an empirical comparison of repeatability and accuracy
Doolittle WF. 1999. Lateral genomics. Trends Cell Biol. 9:M5–M8. between POY and structural alignment. Syst Biol. 56:133–146.
Edgar RC, Batzoglou S. 2006. Multiple sequence alignment. Curr Opin Kluge AG. 1989. A concern for evidence and a phylogenetic hypothesis for
Struct Biol. 16:368–373. relationships among Epicrates (Boidae, Serpentes). Syst Zool. 38:7–25.

Edwards SV. 2008. Is a new and general theory of molecular systematics Knowles LL. 2009. Estimating species trees: methods of phylogenetic
emerging? Evolution. 63:1–19. analysis when there is incongruence across genes. Syst Biol. 58:463–467.

Edwards SV, Liu L, Pearl DK. 2007. High-resolution species trees without Kubatko LS. 2009. Identifying hybridization events in the presence of
concatenation. Proc Natl Acad Sci U S A. 104:5936–5941. coalescence via model selection. Syst Biol. 58:478–488.

Farris JS. 1988. Hennig86, version 1.5. Program and documentation. Port Kubatko LS, Carstens BC, Knowles LL. 2009. STEM: species tree
Jefferson Station (NY): Distributed by the Author. estimation using maximum likelihood for gene trees under coalescence.
Bioinformatics. 25:971–973.
Farris JS, Kallersjo M, Kluge AG, Bult C. 1994. Testing the significance of
incongruence. Cladistics. 10:315–319. Lanyon SM. 1985. Detecting internal inconsistencies in distance data. Syst

Downloaded from https://academic.oup.com/jhered/article/102/1/130/821529 by guest on 27 September 2022

Zool. 38:7–25.
Felsenstein J. 1985. Confidence limits on phylogenies: an approach using
the bootstrap. Evolution. 39:783–791. Larkin MA, Blackshields G, Brown NP, Chenna R, McGettigan PA,
McWilliam H, Valentin F, Wallace IM, Wilm A, Lopez R, et al. 2007. Clustal
Fitch WM, Markowitz E. 1970. An improved method for determining W and Clustal X version 2.0. Bioinformatics. 23:2947–2948.
codon variability in a gene and its application to the rate of fixation of
Lassmann T, Sonnhammer ELL. 2002. Quality assessment of multiple
mutations in evolution. Biochem Genet. 4:579–593.
alignment programs. FEBS Lett. 529:126–130.
Genome 10k Community of Scientists. 2009. Genome 10k: a proposal to obtain
Lee MS, Hugall AF. 2003. Partitioned likelihood support and the evaluation
whole-genome sequence for 10 000 vertebrate species. J Hered. 100:659–674.
of data set conflict. Syst Biol. 52:15–22.
Giribet G. 2005. A review of: ‘‘TNT: tree analysis using New Technology.’’.
Lewis PO. 2001. A likelihood approach to estimating phylogeny from
Syst Biol. 54:176–178.
discrete morphological character data. Syst Biol. 50:913–925.
Giribet G, Wheeler WC. 1999. On gaps. Mol Phylogenet Evol. 13:132–143.
Lewis RL, Beckenbach AT, Mooers AØ. 2005. The phylogeny of the
Goloboff PA. 1994. Nona: A Tree-searching program. Program and subgroups within the melanogaster species group: likelihood tests on COI and
documentation [Internet]. Available from: http://www.zmuc.dk/public COII sequences and a Bayesian estimate of phylogeny. Mol Phylogenet
/phylogeny/Nona-PeeWee. Published by the author, Tucumán, Argentina. Evol. 37:15–24.
Goloboff PA. 1999. Analyzing large datasets in reasonable times: solutions Liu K, Nelesen S, Raghavan S, Linder CR, Warnow T. 2009. Barking up the
for composite optima. Cladistics. 15:415–428. wrong treelength: the impact of gap penalty on alignment and tree accuracy.
Goloboff PA, Farris J, Nixon K. 2003. T.N.T.: Tree Analysis Using New IEEE/ACM Trans Comput Biol Bioinform. 6:7–21.
Technology. Program and documentation [Internet]. Available from: Liu L. 2008. BEST: Bayesian estimation of species trees under the
http://www.zmuc.dk/public/phylogeny/tnt. coalescent model. Bioinformatics. 24:2542–2543.
Goloboff PA, Pol D. 2007. On divide-and-conquer strategies for parsimony Liu L, Pearl DK. 2007. Species trees from gene trees: reconstructing
analysis of large data sets: rec-I-dcm3 vs TNT. Syst Biol. 56:485–495. Bayesian posterior distributions of a species phylogeny using estimated gene
Goloboff PA, Farris JS, Nixon KC. 2008. TNT, a free program for tree distributions. Syst Biol. 56:504–514.
phylogenetic analysis. Cladistics. 24:774–786. Liu L, Yu L, Pearl DK, Edwards SV. 2009. Estimating species phylogenies
Golubchik T, Wise MJ, Easteal S, Jermiin LS. 2007. Mind the gaps: evidence of using coalescence times among sequences. Syst Biol. 58:468–477.
bias in estimates of multiple sequence alignments. Mol Biol Evol. 24:2433–2442. Lopez P, Casane D, Philippe H. 2002. Heterotachy, an important process of
Green PJ. 1995. Reversible jump Markov chain Monte Carlo computation protein evolution. Mol Biol Evol. 19:1–7.
and Bayesian model determination. Biometrika. 82:711–732. Lunter G, Miklós I, Drummond A, Jensen JL, Hein J. 2005. Bayesian
Heled J, Drummond AJ. 2010. Bayesian inference of species trees from coestimation of phylogeny and sequence alignment. BMC Bioinformatics.
multilocus data. Mol Biol Evol. 27:570–580. 6:83.
Hovenkamp P. 2004. Review of TNT—tree analysis using New Maddison WP. 1997. Gene trees in species trees. Syst Biol. 46:523–536.
Technology, ver. 1.0. Cladistics. 20:378–383. Maddison WP, Knowles LL. 2006. Inferring phylogeny despite incomplete
Huelsenbeck JP, Larget B, Alfaro ME. 2004. Bayesian phylogenetic model lineage sorting. Syst Biol. 55:21–30.
selection using reversible jump Markov chain Monte Carlo. Mol Biol Evol. Marshall DC, Simon C, Buckley TR. 2006. Accurate branch length estimation
21:1123–1133. in partitioned Bayesian analyses requires accommodation of among-partition
Kass RE, Raftery AE. 1995. Bayes factors. J Am Stat Assoc. 90:773–795. rate variation and attention to branch length priors. Syst Biol. 55:993–1003.
Katoh K, Kuma K, Toh H, Miyata T. 2005. MAFFT version 5: improvement McCormack JE, Huang H, Knowles LL. 2009. Maximum likelihood
in accuracy of multiple sequence alignment. Nucleic Acids Res. 33:511–518. estimates of species trees: how accuracy of phylogenetic inference depends
upon the divergence history and sampling design. Syst Biol. 58:501–508.
Katoh K, Misawa K, Kuma K, Miyata T. 2002. MAFFT: a novel method
for rapid multiple sequence alignment based on fast Fourier transform. Meier R, Ali F. 2005. The newest kid on the parsimony block: TNT
Nucleic Acids Res. 30:3059–3066. (Tree analysis using new technology). Syst Entomol Am. 30:179–182.
Kawakita A, Sota T, Ascher JS, Ito M, Tanaka H, Kato M. 2003. Evolution Miklós I, Lunter GA, Holmes I. 2004. A ‘‘long indel’’ model for
and phylogenetic utility of alignment gaps within intron sequences of three evolutionary sequence alignment. Mol Biol Evol. 21:529–540.
nuclear genes in bumble bees (Bombus). Mol Biol Evol. 20:87–92. Morgenstern B. 2004. DIALIGN: multiple DNA and protein sequence
Keightley PD, Johnson T. 2004. MCALIGN: stochastic alignment of alignment at BiBiServ. Nucleic Acids Res. 32:W33–W36.
noncoding DNA sequences based on an evolutionary model of sequence Nixon KC. 1999. The parsimony ratchet, a new method for rapid
evolution. Genome Res. 14:442–450. parsimony analysis. Cladistics. 15:407–414.

137
Journal of Heredity 2011:102(1)

Notredame C, Higgins DG, Heringa J. 2000. T-Coffee: a novel method for Simmons MP, Ochoterena H. 2000. Gaps as characters in sequence-based
fast and accurate multiple sequence alignment. J Mol Biol. 302:205–217. phylogenetic analysis. Syst Biol. 49:369–381.
Novák A, Miklós I, Lyngso R, Hein J. 2008. StatAlign: an extendable Simmons MP, Ochoterena H, Carr TG. 2001. Incorporation, relative
software package for joint Bayesian estimation of alignments and homoplasy, and effect of gap characters in sequence-based phylogenetic
evolutionary trees. Bioinformatics. 24:2403–2404. analyses. Syst Biol. 50:454–462.
Nuin PAS, Wang Z, Tillier ERM. 2006. The accuracy of several multiple Stamatakis A. 2006. RAxML-VI-HPC: maximum-likelihood based phylo-
sequence alignment programs for proteins. BMC Bioinformatics. 7:471. genetic analyses with thousands of taxa and mixed models. Bioinformatics.
Nylander JAA, Ronquist F, Huelsenbeck JP, Nieves-Aldrey J. 2004. 22:2688–2690.
Bayesian phylogenetic analysis of combined data. Syst Biol. 53:47–67. Struck TH, Purschke G, Halanych KM. 2006. Phylogeny of Eunicida
Ogden TH, Rosenberg MS. 2007a. Alignment and topological accuracy of (Annelida) and exploring data congruence using a partition addition
the direct optimization approach via POY and traditional phylogenetics via bootstrap alteration (PABA) approach. Syst Biol. 55:1–20.
ClustalW þ PAUP*. Syst Biol. 56:182–193. Suchard MA, Redelings BD. 2006. Bali-Phy: simultaneous Bayesian
Ogden TH, Rosenberg MS. 2007b. How should gaps be treated in inference of alignment and phylogeny. Bioinformatics. 22:2047–2048.

Downloaded from https://academic.oup.com/jhered/article/102/1/130/821529 by guest on 27 September 2022

parsimony? A comparison of approaches using simulation. Mol Phylogenet Swofford DL. 2002. PAUP*: phylogenetic analysis using parsimony (* and
Evol. 42:817–826. other methods), version 4. Sunderland (MA): Sinauer Associates.
Page RDM, Charleston MA. 1997. From gene to organismal phylogeny: Swofford DL, Olsen GJ, Waddell PJ, Hillis DM. 1996. Phylogenetic
reconciled trees and the gene tree species tree problem. Mol Phylogenet inference. In: Hillis DM, Moritz C, Mable BK, editors. Molecular
Evol. 7:231–240. systematics. 2nd ed. Sunderland (MA): Sinauer Associates. p. 407–514.
Pagel M, Meade A. 2004. A phylogenetic mixture model for detecting pattern- Talavera G, Castresana J. 2007. Improvement of phylogenies after
heterogeneity in gene sequence or character-state data. Syst Biol. 53:571–581. removing divergent and ambiguously aligned blocks from protein sequence
Pagel M, Meade A. 2005. Mixture models in phylogenetic inference. In: alignments. Syst Biol. 56:564–577.
Gascuel O, editor. Mathematics of evolution and phylogeny. Oxford: Thompson JD, Higgins DG, Gibson TJ. 1994. CLUSTAL W: improving
Clarendon Press. p. 121–142. the sensitivity of progressive multiple sequence alignment through sequence
Pagel M, Meade A. 2008. Modelling heterotachy in phylogenetic inference weighting, position-specific gap penalties and weight matrix choice. Nucleic
by reversible-jump Markov chain Monte Carlo. Philos Trans R Soc B Biol Acids Res. 22:4673–4680.
Sci. 363:3955–3964. Thorne JL, Kishino H, Felsenstein J. 1991. An evolutionary model for
Pamilo P, Nei M. 1988. Relationships between gene trees and species trees. maximum likelihood alignment of DNA sequences. J Mol Evol.
Mol Biol Evol. 5:568–583. 33:114–124.

Philippe H, Zhou Y, Brinkmann H, Rodrigue N, Delsuc F. 2005. Heterotachy Thorne JL, Kishino H, Felsenstein J. 1992. Inching towards reality—an
and long-branch attraction in phylogenetics. BMC Evol Biol. 5:50. improved likelihood model of sequence evolution. J Mol Evol. 34:3–16.

Raftery AE. 1996. Hypothesis testing and model selection. In: Gilks WR, Varón A, Vinh LS, Wheeler WC. 2009. POY version 4: phylogenetic
Richardson S, Spiegelhalter DJ, editors. Markov chain Monte Carlo in analysis using dynamic homologies. Cladistics. 26:72–85.
practice. London: Chapman Hall. p. 163–188. Venditti C, Meade A, Pagel M. 2008. Phylogenetic mixture models can
Redelings BD, Suchard MA. 2005. Joint Bayesian estimation of alignment reduce node-density artifacts. Syst Biol. 57:286–293.
and phylogeny. Syst Biol. 54:401–418. Wallace IM, Blackshields G, Higgins DG. 2005. Multiple sequence
Redelings BD, Suchard MA. 2007. Incorporating indel information into alignments. Curr Opin Struct Biol. 15:261–266.
phylogeny estimation for rapidly emerging pathogens. BMC Evol Biol. 7:40. Wallace IM, O’Sullivan O, Higgins DG, Notredame C. 2006. M-Coffee:
Rieseberg LH, Baird SJ, Gardner KA. 2000. Hybridization, introgression, combining multiple sequence alignment methods with T-Coffee. Nucleic
and linkage evolution. Plant Mol Biol. 42:205–224. Acids Res. 34:1692–1699.

Roberts TE, Sargis EJ, Olson LE. 2009. Networks, trees, and treeshrews: Wheeler W. 1996. Optimization alignment: the end of multiple sequence
assessing support and identifying conflict with multiple loci and alignment in phylogenetics. Cladistics. 12:1–9.
a problematic root. Syst Biol. 58:257–270. Yang Z. 1994. Maximum likelihood phylogenetic estimation from DNA
Rokas A, Williams BL, King N, Carroll SB. 2003. Genome-scale approaches sequences with variable rates over sites: approximate methods. J Mol Evol.
to resolving incongruence in molecular phylogenies. Nature. 425:798–804. 39:306–314.

Ronquist F, Huelsenbeck JP. 2003. MrBayes 3: Bayesian phylogenetic Young ND, Healy J. 2003. GapCoder automates the use of indel characters
inference under mixed models. Bioinformatics. 19:1572–1574. in phylogenetic analysis. BMC Bioinformatics. 4:6.

Schulte JAII, Cartwright EM. 2009. Phylogenetic relationships among Zaldivar-Riverón A, Nieto-Montes de Oca A, Manrı́quez-Morán N,
iguanian lizards using alternative partitioning methods and TSHZ1: a new Reeder TW. 2008. Phylogenetic affinities of the rare and enigmatic limb-
phylogenetic marker for reptiles. Mol Phylogenet Evol. 50:391–396. reduced Anelytropsis (Reptilia: Squamata) as inferred with mitochondrial
16S rRNA sequence data. J Herpetol. 42:303–311.
Siddall ME. 2001. Philosophy and phylogenetic inference: a comparison of
likelihood and parsimony methods in the context of Karl Popper’s writings Zwickl DJ. 2006. Genetic algorithm approaches for the phylogenetic
on corroboration. Cladistics. 17:395–399. analysis of large biological sequence datasets under the maximum likelihood
criterion [PhD dissertation]. [Austin (TX)]: The University of Texas at
Siddall ME, Kluge AG. 1997. Probabilism and phylogenetic inference.
Austin.
Cladistics. 13:313–336.
Simmons MP. 2004. Independence of alignment and tree search. Received April 21, 2010; Revised June 11, 2010;
Mol Phylogenet Evol. 31:874–879. Accepted July 12, 2010
Simmons MP, Müller K, Norton AP. 2007. The relative performance of
indel-coding methods in simulation. Mol Phylogenet Evol. 44:724–740. Corresponding Editor: William Murphy

138