MEGA5: Molecular Evolutionary Genetics Analysis Using
Maximum Likelihood, Evolutionary Distance, and Maximum
Koichiro Tamura,1,2Daniel Peterson,2Nicholas Peterson,2Glen Stecher,2Masatoshi Nei,3and
1Department of Biological Sciences, Tokyo Metropolitan University, Hachioji, Tokyo, Japan
2Center for Evolutionary Medicine and Informatics, The Biodesign Institute, Arizona State University
3Department of Biology and the Institute of Molecular Evolutionary Genetics, The Pennsylvania State University
4School of Life Sciences, Arizona State University
*Corresponding author: E-mail: email@example.com.
Associate editor: Naoki Takebayashi
Comparative analysis of molecular sequence data is essential for reconstructing the evolutionary histories of species and
inferring the nature and extent of selective forces shaping the evolution of genes and species. Here, we announce the
release of Molecular Evolutionary Genetics Analysis version 5 (MEGA5), which is a user-friendly software for mining online
databases, building sequence alignments and phylogenetic trees, and using methods of evolutionary bioinformatics in basic
biology, biomedicine, and evolution. The newest addition in MEGA5 is a collection of maximum likelihood (ML) analyses
for inferring evolutionary trees, selecting best-fit substitution models (nucleotide or amino acid), inferring ancestral states
and sequences (along with probabilities), and estimating evolutionary rates site-by-site. In computer simulation analyses,
ML tree inference algorithms in MEGA5 compared favorably with other software packages in terms of computational
efficiency and the accuracy of the estimates of phylogenetic trees, substitution parameters, and rate variation among sites.
The MEGA user interface has now been enhanced to be activity driven to make it easier for the use of both beginners and
experienced scientists. This version of MEGA is intended for the Windows platform, and it has been configured for
effective use on Mac OS X and Linux desktops. It is available free of charge from http://www.megasoftware.net.
The Molecular Evolutionary Genetics Analysis (MEGA) soft-
tric, integrated suite of tools for statistical analyses of DNA
and protein sequence data from an evolutionary standpoint.
Over the years, it has grown to include tools for sequence
alignment,phylogenetic tree reconstruction and visualization,
testing an array of evolutionary hypotheses, estimating se-
quence divergences, web-based acquisition of sequence data,
and expert systems to generate natural language descriptions
of the analysis methods and data chosen by the user (Kumar
et al. 1994, 2008; Kumar and Dudley 2007). With the fifth
major release, the collection of analysis tools in MEGA
has now broadened to include the maximum likelihood
(ML) methods for molecular evolutionary analysis. Table 1
contains a summary of all statistical methods and models
the following, we provide a brief description of methodo-
logical advancements, along with relevant research results,
and technical enhancements in MEGA5.
Model Selection for Nucleotide and Amino
MEGA5 now contains facilities to evaluate the fit of major
models of nucleotide and amino acid substitutions, which
are frequently desired by researchers (Posada and Crandall
1998; Nei and Kumar 2000; Yang 2006) (fig. 1A). For nucle-
otide substitutions, the GTR and five nested models are
available, whereas six models with and without empirical
frequencies(þF)have been programmedfor the aminoacid
substitutions (Table 1). MEGA5 provides the goodness-
of-fit (see below) of the substitution models with and
without assuming the existence of evolutionary rate vari-
ation among sites, which is modeled by a discrete Gamma
distribution (þG) (Yang 1994) and/or an allowance for the
presence of invariant sites (þI) (Fitch and Margoliash 1967;
Fitch 1986; Shoemaker and Fitch 1989). This results in an
evaluation of 24 and 48 models for nucleotide and amino
acid substitutions, respectively. For each of these models,
MEGA5 provides the estimated values of shape parameter
of the Gamma distribution (a), the proportion of invariant
sites, and the substitution rates between bases or residues,
as applicable. Depending on the model, the assumed or
observed values of the base or amino acid frequencies used
in the analysis are also provided. This information enables
researchers to quickly examine the robustness of the esti-
mates of evolutionary parameters under different models
of substitutions and assumptions about the distribution of
of each model to the data is measured by the Bayesian
information criterion (BIC, Schwarz 1978) and corrected
© The Author 2011. Published by Oxford University Press on behalf of the Society for Molecular Biology and Evolution. All rights reserved. For permissions, please
Mol. Biol. Evol. 28(10):2731–2739. 2011doi:10.1093/molbev/msr121Advance Access publication May 4, 20112731
Akaike information criterion (AICc, Hurvich and Tsai 1989)
(see also Posada and Buckley 2004). By default, MEGA5 lists
models with decreasing BIC values (see below for the rea-
son and caveats), along with log likelihood as well as AICc
values for each model.
In the ML methods for evaluating the fit of substitution
models to the data, an evolutionary tree is needed. MEGA5
automatically infers the evolutionary tree by the Neighbor-
Joining (NJ) algorithm that uses a matrix of pairwise distan-
ces estimated under the Jones–Thornton–Taylor (JTT)
model for amino acid sequences or the Tamura and Nei
(1993) model for nucleotide sequences (Saitou and Nei
1987; Jones et al. 1992; Tamura and Nei 1993; Tamura
et al. 2004). Branch lengths and substitution rate parame-
may provide their own tree topology in the Newick (New
Hampshire) format for use in this model selection (fig. 1B).
However, the automatic option is expected to be fre-
quently used because trees are rarely known a priori.
trees in MEGA5 on the process of model selection by com-
puter simulation. These simulations used 448 sets of evo-
lutionary parameters (base frequencies, sequence length,
mean evolutionary rate, and transition–transversion rate
ratio) derived from real sequence data (see Rosenberg
and Kumar 2001) and introduced four different levels of
rate variation among sites for each parameter set (Gamma
shape parameter, a 5 0.25, 0.5, 1.0, and 2.0). Results
showed that the best-fit models produced by using auto-
matically generated trees were the same as those inferred
using the true tree for ?93% of the data sets according to
the BIC and AICc criteria (fig. 2A).
For an overwhelming majority of data sets, AICc selected
the most complex model (see also Ripplinger and Sullivan
2008). But, both BIC and AICc selected substitutions models
that were more complex than the true model (Posada and
was among the top-3 when BIC was used and among the
top-5 when AICc was used. When the rate variation among
sites (þI) along with discrete gamma rate categories (þG)
were favored for virtually every data set. This means that
a discrete gamma (þG) model using a small number of cat-
egories (4), which is a common practice, coupled with an
allowance for invariant sites (þI) is better at approximating
the continuous Gamma distribution used in the simulation
when the rate variation among sites is severe. This was con-
firmed by comparing the ML value for the fit of HKY þ G
model (10 categories) with the ML value for GTR þ G þ I
model using only four discrete gamma categories. The for-
mer performed slightly better than the latter, even though
the latter involved a more complex model.
BICselectsthe true modelfor .70%of the datasets. Incon-
trast, AICc selects the correct model only 35% of the time.
Therefore, we rank the models by BIC in MEGA5 (fig. 1C).
However, the choiceof criterion toselectthe best-fit models
is rather complicated, and researchers should explore model
selection based on AICc values and other available methods
for evolutionary analyses in which choice of model is known
ysis to select the best model, MEGA5 provides exporting of
results in Microsoft Excel/Open Office and comma sepa-
rated values formats.
These simulation results also provided us with an
opportunity to evaluate the estimates of a obtained by us-
ing the automatically generated tree and to compare them
to those obtained by using the true tree under the correct
model of substitution. The means and standard deviations
of these estimates were very similar to the true values and
virtually identical for automatically generated and true
trees (fig. 2B). Similarly, the overall estimates of
Table 1. A Summary of Analyses and Substitution Models in MEGA5
DNA, codon, and protein alignments; both manual and automated alignments with trace file Editor. Built-in automated aligners: CLUSTALW
Major analyses (statistical approach in parentheses)
Models and parameters: Select Best-Fit Substitution Model* (ML); test pattern homogeneity; Estimate Substitution Pattern (MCL, ML*);
Estimate Rate Variation Among Sites* (ML); Estimate Transition/Transversion Bias (MCL, ML*); Estimate Site-by-Site Rates* (ML).
Infer phylogenies: Infer Phylogenetic Trees (NJ, ML*, ME, MP); Phylogeny Tests (Bootstrap and Branch-length tests); Branch-and-Bound Exact
Search (MP); Heuristic Searches: Nearest-Neighbor-Interchange (NNI; ML*, ME, MP), Close-Neighbor-Interchange (CNI; ML*, ME, MP), and
Compute distances: Pairwise and Diversity; Within- and Between-Group Distances; Bootstrap and Analytical Variances; separate distances by
Site Degeneracy, Codon Sites; Separation of Distances in Transitions and Transversions; Separate Nonsynonymous and Synonymous Changes
Tests of Selection: For Complete Sequences or Set of Codons; Sequence Pairs or Groups (Within and Between)
Ancestral Sequences: Infer by ML with Relative Probabilities for bases or residues* or by MP (all parsimonious pathways)
Molecular Clocks: Tajima’s 3-Sequence Clock Test*; Likelihood Ratio Test (ML) for a Topology*; Estimate Branch Lengths under Clock*
Substitution models (1F 5 with empirical frequencies; REV 5 reversible)
DNA: General Time Reversible (GTR)*, Tamura–Nei, Hasegawa–Kishino–Yano*, Tamura Three-Parameter, Kimura Two-Parameter, Tajima–
Codons: Nei–Gojobori (original and modified), Li–Wu–Lou (original and modified)
Protein: Poisson, Equal-Input, Dayhoff (1F), Jones–Taylor–Thornton (1F), Whelan and Goldman (1F)*, Mitochondrial REV (1F)*, Chloroplast
REV (1F)*, Reverse Transcriptase REV (1F)*
Rate Variation and Base Compositions: Gamma rates (G) and Invariant sites (I)* models; Incorporate Compositional Heterogeneity.
NOTE.—MCL, maximum composite likelihood; ME, minimum evolution; MP, maximum parsimony. An asterix (*) denotes features that are new in MEGA5.
Tamura et al. · doi:10.1093/molbev/msr121
transition–transversion ratio (R) were close to the true
value for both automatically generated and true trees
with MEGA5 is useful, as a first approximation, in estimat-
ing evolutionary substitution parameters and evaluating
relative fits of models.
Inferring ML Trees
MEGA5 now provides the ML method to infer evolutionary
trees and conduct the bootstrap test for nucleotide and
amino acid alignments (Felsenstein 1981, 1985). Because
the ML method is computationally demanding, we provide
heuristic methods that search for the MLtreeby topological
rearrangements of an initial tree (Swofford 1998; Nei and
Kumar 2000; Guindon and Gascuel 2003; Stamatakis et al.
2005). The initial tree for the ML search can be supplied
by the user (Newick format) or generated automatically
by applying NJ and BIONJ algorithms to a matrix of pairwise
distances estimated using a maximum composite likelihood
approach for nucleotide sequences and a JTT model for
amino acid sequences (Saitou and Nei 1987; Jones et al.
1992; Gascuel 1997; Tamura et al. 2004). For the user-
selected data subset that contains sites with insertion–
deletions and missing data, we begin by temporarily
obtaining a site coverage parameter such that the number
of ambiguous states and insertion–deletions per sequence
are the lowest. This site coverage parameter is then used
togeneratea data subsetfor estimating evolutionary distan-
ces to build an initial tree along with branch lengths. We
FIG. 1. EvaluatingthefitofsubstitutionmodelsinMEGA5.(A)The‘‘Models’’menuinthe‘‘ActionBar’’providesaccesstothefacility.(B)An‘‘Analysis
gaps. In addition to the ‘‘Complete Deletion’’ and ‘‘Pairwise Deletion’’ options, MEGA5 now includes a ‘‘Partial Deletion’’ option that enables users to
an alignmentgap, missingdatum, orambiguous base/aminoacid. For proteincoding nucleotidesequences, userscan chooseto analyzenucleotideor
fits, number of parameters (branch lengthsþ model parameters), and estimates of evolutionary parameters for Drosophila Adhsequence data which
by BIC), data subset selected, and the analysis option chosen. This figure is available in color online and in black and white in print.
MEGA5: Molecular Evolutionary Genetics Analysis · doi:10.1093/molbev/msr121
found this approach to produce better initial estimates
when there are many insertions–deletions and missing data
subset is restored and used in all subsequent calculations.
the initial tree, such that the alternative trees differ in one
branching pattern. One can expand the search space by
using the CNI option in which alternative trees with
all the alternative trees produced by the branch swapping
are made simultaneously. If several single rearrangements
are found to improve ML values for any branch, we choose
the rearrangement that leads to the highest improvement
in the ML value. We do not skip any branch swaps as long
as it improves the ML value. In order to make major com-
putational time savings, we do skip the evaluation of alter-
native topologies generated by rearrangements involving
branches whose lengths are more than three times longer
than their approximate standard errors. We use the second
derivative of the ML score to generate approximate stan-
dard errors (Edwards 1972) during the branch length opti-
mizations. Therefore, starting with systematic topological
rearrangements of the initial tree, we discover trees with
a higher ML value. These trees are subjected to new rounds
of rearrangements, and this iterative process continues
until no trees with greater likelihood can be found.
We tested the performance (time and accuracy) of the
NNI and CNI searches in MEGA5 by means of
computer-simulated data sets containing 66 sequences
(see Materials and Methods). We compared the time taken
to complete these heuristic searches with each other and
with those needed by PhyML version 3.0 (Guindon et al.
2010) and RaxML version 7.0 (Stamatakis 2006). Results
showed that, on average, a CNI search requires twice
the time of an NNI search in MEGA5 (fig. 3A). Speeds
of the MEGA5-NNI and MEGA5-CNI searches were similar
to RaxML7-Mix and RaxML7-G, respectively. But, they were
tively (fig. 3A). Similar trends were observed for another
simulated dataset in which an increasingly larger number
of sequences were analyzed (fig. 3B; 20–765 sequence data
sets). For these data, the ML heuristic time increase shows
a power trend with the increasing number of sequences
(fig. 3B). It is important to note that the RaxML will be fast-
er than MEGA5 if the user’s machine is equipped with mul-
tiple processor and/or multicore CPUs because parallel
versions of MEGA5 are yet to be implemented.
Even though different programs and search options show
of the inferred ML trees were found to be rather similar. The
66 sequences (fig. 4A) and 765 sequences (fig. 4B). Therefore,
ML methods in MEGA5 appear to be comparable to other
widely used ML implementations in terms of computational
time and phylogenetic accuracy. In these simulations, we also
FIG. 2. Comparison of the best-fit model identified by using
automatically generated and true trees for 1,792 computer
simulated 66-sequence data sets. (A) The percentage of datasets
for which the use of an automatically generated tree produces
the same best-fit model as does the use of the true tree. Results
are shown from datasets simulated with four different values of
the gamma parameter (a) for rate variation among sites. (B) The
estimates of a when using the automatically generated trees
(filled bars) and the true tree (open bars). The average a and ±1
standard deviation are depicted on each bar; 10 discrete Gamma
categories were used. (C) The relationship of true and estimated
transition–transversion ratio, R, when using automatically
generated trees for data simulated with a 5 0.25. The value of
R becomes 0.5 when the transition–transversion rate ratio, j, is
1.0 in Kimura’s two-parameter model. The slope of the linear
regression was 1.005, with the intercept passing through the
origin (r25 0.98). Using the true tree, slope and r2values were
1.007 and 0.98, respectively. The absolute average difference
between the two sets of estimates was 0.2% (maximum
difference 5 5.2%). Similar results were obtained for data
simulated with a 5 0.5, 1.0, and 2.0.
Tamura et al. · doi:10.1093/molbev/msr121
over all simulation cases.
Inference of Ancestral States and Sequences
MEGA5 now provides inferences of ancestral states and
sequences using the empirical Bayesian method (fig. 5).
Given a phylogenetic tree, branch lengths are estimated
under a user-selected model of nucleotide or amino acid
erated for each possible ancestral state assignment for each
node (Yang et al. 1995). With this addition, users can now
explore ancestral sequences inferred using maximum parsi-
mony and ML methods in MEGA5. However, the latter is
among multiple equally likely (most parsimonious) assign-
ments by using the posterior probabilities for each possible
nucleotide or amino acid assignment. Furthermore, it is
have undergone multiple substitutions over the whole tree
ancestral sequences. However, note that the reconstructed
ancestral sequences are not real observed data and may
involve systematic biases and random errors, especially for
if they are to be used in further statistical analysis.
Position-by-Position Evolutionary Rates
For both nucleotide and amino acid sequence data, users
can estimate relative rates of molecular evolution position-
by-position in MEGA5. Users select the number of discrete
categories to approximate the Gamma distribution, specify
whether or not to model invariant positions, and choose
a nucleotide or amino acid substitution model. As men-
tioned earlier, they can use an automatically generated
0 50 100 150 200
No. of Sequences
200 300 400
No. of Sequences (x)
500 600 700 800
0 100 200
300 400 500 600
FIG. 3. Comparison of the computational speed of ML heuristic
RaxML7 (G and MIX), and PhyML3 (NNI and SPR) heuristic searches
for 1,792 simulated data sets containing 66 sequences each. Bars are
shown with ±1 standard deviation. Three data sets were excluded
showing the time taken to search for the ML tree for alignments that
contain 20–200 and 200–765 sequences of 2,000 base pairs. The
discrete categories for the Gamma distribution and a GTR model of
nucleotide substitution (see Materials and Methods for simulation
procedures, analysis descriptions, and computer hardware used). G,
GTRGAMMA with four discrete Gamma categories; MIX, mixed
method of using CAT and GAMMA models.
50% 60% 70% 80% 90% 100%
50% 60% 70% 80% 90%
FIG. 4. Accuracies of heuristic ML trees produced by MEGA5,
RaxML7, and PhyML3 programs. Shown are the proportions of
interior branches (tree partitions) inferred correctly, along with ±1
standard deviation, for simulated data sets containing (A) 66
sequences and (B) 765 sequences. G, GTRGAMMA with four
discrete Gamma categories; MIX, mixed method of using CAT and
MEGA5: Molecular Evolutionary Genetics Analysis · doi:10.1093/molbev/msr121
topology, but it should be done carefully because the site-
specific estimates of the evolutionary rate may depend on
the evolutionary tree used (see Mayrose et al. 2005). No
information on sequence divergence times is needed for
estimating relative rates of evolution over sites, where all
individual relative rates are scaled such that the average
relative rate over all positions is equal to 1. This means that
positions showing a relative rate less than 1 are inferred to
be more highly conserved than the average conservation of
sites in the alignment. Whenever available, these results are
automatically exported directly to statistical analysis soft-
ware, including Microsoft Excel, which can be used to gen-
erate sequence-wide profiles and conduct further analyses.
ML Molecular Clocks and Linearized Tree
In addition to Tajima’s nonparametric test of molecular
clock for three sequences (Tajima 1993), we have now
added a likelihood ratio test of the molecular clock where
the ML value for a given tree assuming the rate uniformity
tion. In the output, primary information along with the
P value of rejecting the null hypothesis of equal rates under
a v2distribution is presented. This test is expected to reject
the null hypothesis when applied to data sets containing
many sequences or long sequences as the strict equality of
evolutionary rates among lineages is frequently violated.
On the other hand, the estimates of branch lengths, and
thus interior node depths, in a tree obtained under the as-
sumption of a molecular clock can be useful to generate
a rough idea about the relative timing of sequence
divergence events (e.g., Takezaki et al. 1995). Of course,
such estimates should be used cautiously.
lengths by assuming equal evolutionary rate among line-
ages. With this addition, users can now produce linearized
trees using pairwise distances as well as the ML method.
One can manually calibrate the molecular clock by setting
the divergence time for any one node in the tree, which
produces divergence times for all other nodes in the tree.
confidence intervals from the variance of the node height
computed usingthecurvature method(e.g., Schrago 2006).
Note that this procedure may underestimate the variance
stancy. The estimated node heights may be biased because
of this reason, as well. So, the confidence intervals
presented are not appropriate for hypothesis testing.
We have also introduced many improvements to enhance
MEGA’s usability. First, MEGA5’s central user interface has
now become activity driven where a launch bar provides
direct access to the growing suite of tools according to
the type of analysis needed through the ‘‘Action Bar’’
(fig. 6). Once a user selects what they wish to compute,
MEGA5 prompts for a data file to use and the methods
and data subsets to employ. This wizard-style layout will
make MEGA5 easier for beginners and expert users alike.
In this spirit, we have now added native support for the
widely used FASTA file format for sequence data, and se-
quence data can now be aligned using the MUSCLE soft-
ware, which is very fast and accurate for data sets
containing a large number of sequences (Edgar 2004). Be-
cause MEGA now accepts user trees for heuristic searches,
for molecular clock tests, and for ancestral sequence recon-
struction, we have included a tree topology editor that is
useful for creating trees and editing existing topologies by
using drag-and-drop of branches.
Operating Systems and Platforms
In a recent survey oflong-term MEGAusers, wehave found
that both Mac OS X and Linux platforms are used by a sub-
stantial number of researchers (one out of four). Therefore,
we havebeen optimizing the use ofMEGA5 on the Mac OS
X and Linux platforms. For Mac OS X, we have now devel-
software so that the installation of MEGA5 is as simple as
installing native Mac applications. WINE is a translation
layer capable of running native Windows applications
on POSIX compatible operating systems, such as Mac
OS and Linux, and has two major advantages over using
an emulation layer (i.e., virtualization software). First, by
not using virtualization, users are not required to purchase
a license for an additional operating system. Second, instal-
lation is simplified as there is no need to create and/or in-
stall an operating system disk image. As a result, Mac OS X
users are able to use MEGA5 as seamlessly as if they were
operating it on the Windows platform for which MEGA5
was originally developed. Similar provisions have been
made for the use of MEGA5 on the Linux platform. In
our tests, we found that calculations in MEGA5 on
Mac OS X and Linux are ,5% slower than Windows. This
difference is rather small because all calculations via
WINE are executed directly on the CPU like any other
FIG. 5. Position-specific inferred ancestral states in a primate opsin
phylogeny and the posterior probabilities of alternative amino acids
at that position. See MEGA5 Examples directory for the data file and
Nei and Kumar (2000, p. 212–213) for a description of the data. This
figure is available in color online and in black and white in print.
Tamura et al. · doi:10.1093/molbev/msr121
native application in Mac OS X and Linux. In contrast, the
MEGA5 user interface is rendered via emulation by WINE,
which can sometimes result in a slowdown when drawing
on the screen. But, this is becoming less noticeable with
contemporary CPUs that are extremely fast. This en-
hancement is likely to make MEGA5 more useable for
a greater number of researchers.
In summary, MEGA5 now provides analysis tools for three
major types (ML, MP, and evolutionary distances) of statis-
tical methods of molecular evolution (table 1 and fig. 6).
These facilities not only make MEGA useful for more re-
searchers but also enable researchers to evaluate the robust-
ness of their results by comparing inferences from multiple
methods under a variety of statistical models. In the future,
we will continue to develop MEGA with a focus on imple-
menting faster algorithms for phylogenetic inference, inte-
grating more third party tools, and upgrading the
effectively. As always, all versions of MEGA are available free
of charge from http://www.megasoftware.net.
Materials and Methods
We generatedtwosets of nucleotide sequencedata byusing
phylogenetic relationships among mammals was used (see
independent set of evolutionary parameter values (base fre-
sition–transversion rate ratio) estimated from the real
sequence data (Rosenberg and Kumar 2003). For each
lengths of the model tree were estimated using the corre-
ated under the Hasegawa–Kishino–Yano (HKY) (Hasegawa
egories. This resulted in a total of 1,792 alignment sets.
We also generated DNA sequence alignments containing
20–765 taxa, which were based on the corresponding sized
trees derived from a master phylogeny of 765 taxa (see
supplementary fig. S1, Supplementary Material online, in
Battistuzzi et al. ). This master phylogeny was
obtained by pruning taxa and groups from the tree of
1,671 families in the Timetree of Life (Hedges and Kumar
2009), such that the final tree was strictly bifurcating.
The resultant tree of 765 taxa was scaled to time and
spanned 4.2 billion years of evolution. This master topology
was subsampled to produce model trees used to generate
the sequence alignments containing varying number of taxa
(20, 40, 60, ..., 500), with one set containing all 765 taxa.
Sequences were simulated using SeqGen (Rambaut and
Grassly 1997) under the HKY (Hasegawa et al. 1985) model
of nucleotide substitution with a G þ C content of 48% and
a transition–transversion rate ratio of 1.05, which were esti-
mated from an alignment of small subunit rRNA sequences
from 800 taxa of animals, fungi,plants, and archaebacteria.In
order to make the evolutionary rate heterogeneous among
tip and internal lineages, rates were varied randomly by
drawing them from a uniform distribution with boundaries
±5% of the expected rate in each branch independently. We
used substitution rates of 0.025, 0.050, and 0.100 per base
pair per billion years to establish branch lengths. In total,
530 data sets were generated in this way and the results
are presented in the main text. We also conducted
290–765 taxa simulations in which sequences evolved four
times faster (0.4 substitutions per site per billion years) and
found the differences between methods were very similar to
those reported here (results not shown).
A benchmark comparison of ML phylogenetic inference
between MEGA5, RAxML7, and PhyML3 was performed for
all simulated datasets by collecting the computational and
phylogenetic performance of these programs, including
FIG. 6. The MEGA5 ‘‘Action Bar’’ and associated action menus. This figure is available in color online and in black and white in print.
MEGA5: Molecular Evolutionary Genetics Analysis · doi:10.1093/molbev/msr121
execution time (in seconds), the estimate of a gamma
shape parameter, ML values, and topological accuracy. Be-
cause Windows is MEGA5’s native operating system, Win-
dows executables were used for PhyML (version 3.0) and
RAxML (version 7.04). All analyses were conducted on
computers with identical hardware (Intel Q8400 2.66
GHz Quad Core processor and 6 GB RAM) and operating
systems (64-bit Windows 7 Enterprise Edition). For direct
comparison, each program was executed serially in a single
thread of execution with one core utilized per dataset.
In order to generate comparable results on time and
accuracy, we used identical substitution models and dis-
crete gamma options across all programs. Because the fast-
est heuristic search in RaxML, MIX, assumes a general time
reversible (GTR) model with four discrete gamma rate
categories, we used these options in all cases, unless noted
otherwise. For all three programs, analyses were conducted
using theautomatically generated initial trees and selecting
the default options. And, heuristic searches starting with
the initial trees were conducted with two different levels
of branch rearrangements: quick searches (Nearest Neigh-
bor Interchange [NNI] for MEGA5 and PhyML and MIX for
RaxML) and slow searches (Close Neighbor Interchange
[CNI] for MEGA5, Subtree–Pruning–Regrafting [SPR] for
PhyML, and GTRGAMMA for RaxML). The accuracy of
phylogenetic tree of n taxa was estimated from the topo-
logical distance (dT) between the inferred tree and the true
topology was given by (n ? 3 ? ½dT)/(n ? 3).
We thank the colleagues, students, and volunteers who
spent countless hours testing the early release versions
of MEGA5. Many facets of the user-interface design
benefited from the extensive comments of the members
of our laboratories and of users at large. We also thank
Mr PaulBilling-Ross for his help with computer simulations
and Ms Carol Williams for editorial support. The MEGA
software project is supported by research grants from Na-
tional Institutes of Health to S.K. and M.N.
Alfaro ME, Huelsenbeck JP. 2006. Comparative performance of
Bayesian and AIC-based measures of phylogenetic model
uncertainty. Syst Biol. 55:89–96.
Battistuzzi FU, Billing-Ross P, Paliwal A, Kumar S. 2011. Fast and slow
implementations of relaxed clock methods show similar patterns
of accuracy in estimating divergence times. Mol Biol Evol.
Edgar RC. 2004. MUSCLE: a multiple sequence alignment method with
reduced time and space complexity. BMC Bioinformatics. 5:113.
Edwards AWF. 1972. Likelihood; an account of the statistical
concept of likelihood and its application to scientific inference.
Cambridge (UK): Cambridge University Press.
Felsenstein J. 1981. Evolutionary trees from DNA sequences:
a maximum likelihood approach. J Mol Evol. 17:368–376.
Felsenstein J. 1985. Confidence limits on phylogenies: an approach
using the bootstrap. Evolution 39:783–791.
Fitch WM. 1986. An estimation of the number of invariable sites is
necessary for the accurate estimation of the number of
nucleotide substitutions since a common ancestor. Prog Clin
Biol Res. 218:149–159.
Fitch WM, Margoliash E. 1967. A method for estimating the number
of invariant amino acid coding positions in a gene using
cytochrome c as a model case. Biochem Genet. 1:65–71.
Gascuel O. 1997. BIONJ: an improved version of the NJ algorithm
based on a simple model of sequence data. Mol Biol Evol.
Guindon S, Dufayard JF, Lefort V, Anisimova M, Hordijk W,
Gascuel O. 2010. New algorithms and methods to estimate
maximum-likelihood phylogenies: assessing the performance of
PhyML 3.0. Syst Biol. 59:307–321.
Guindon S, Gascuel O. 2003. A simple, fast, and accurate algorithm
to estimate large phylogenies by maximum likelihood. Syst Biol.
Hasegawa M, Kishino H, Yano T. 1985. Dating of the human-ape
splitting by a molecular clock of mitochondrial DNA . J Mol Evol.
Hedges SB, Kumar S. 2009. Discovering the timetree of life. In:
SB Hedges, S Kumar, editors. The timetree of life. New York:
Oxford University Press. p. 3–18.
Hurvich CM, Tsai C-L. 1989. Regression and time series model
selection in small samples. Biometrika 76:297–307.
Jones DT, Taylor WR, Thornton JM. 1992. The rapid generation of
mutation data matrices from protein sequences. Comput Appl
Kumar S, Dudley J. 2007. Bioinformatics software for biologists in the
genomics era. Bioinformatics 23:1713–1717.
Kumar S, Nei M, Dudley J, Tamura K. 2008. MEGA: a biologist-
centric software for evolutionary analysis of DNA and protein
sequences. Brief Bioinform. 9:299–306.
Kumar S, Tamura K, Nei M. 1994. MEGA: molecular evolutionary
genetics analysis software for microcomputers. Comput Appl
Mayrose I, Mitchell A, Pupko T. 2005. Site-specific evolutionary rate
inference: taking phylogenetic uncertainty into account. J Mol
Nei M, Kumar S. 2000. Molecular evolution and phylogenetics.
Oxford: Oxford University Press.
Posada D, Buckley TR. 2004. Model selection and model averaging in
phylogenetics: advantages of Akaike information criterion and
Bayesian approaches over likelihood ratio tests. Syst Biol.
Posada D, Crandall KA. 1998. MODELTEST: testing the model of
DNA substitution. Bioinformatics 14:817–818.
Rambaut A, Grassly NC. 1997. Seq-Gen: an application for the
Monte Carlo simulation of DNA sequence evolution along
phylogenetic trees. Comput Appl Biosci. 13:235–238.
Ripplinger J, Sullivan J. 2008. Does choice in model selection affect
maximum likelihood analysis? Syst Biol. 57:76–85.
Ripplinger J, Sullivan J. 2010. Assessment of substitution model
adequacy using frequentist and Bayesian methods. Mol Biol Evol.
Rosenberg MS, Kumar S. 2001. Incomplete taxon sampling is not
a problem for phylogenetic inference. Proc Natl Acad Sci U S A.
Rosenberg MS, Kumar S. 2003. Heterogeneity of nucleotide
frequencies among evolutionary lineages and phylogenetic
inference. Mol Biol Evol. 20:610–621.
Saitou N, Nei M. 1987. The Neighbor-Joining Method—a new method
for reconstructing phylogenetic trees. Mol Biol Evol. 4:406–425.
Schrago CG. 2006. An empirical examination of the standard errors
of maximum likelihood phylogenetic parameters under the
molecular clock via bootstrapping. Genet Mol Res. 5:233–241.
Schwarz G. 1978. Estimating the dimension of a model. Ann Stat.
Tamura et al. · doi:10.1093/molbev/msr121
Shoemaker JS, Fitch WM. 1989. Evidence from nuclear sequences
that invariable sites should be considered when sequence
divergence is calculated. Mol Biol Evol. 6:270–289.
Stamatakis A. 2006. RAxML-VI-HPC: maximum likelihood-based
phylogenetic analyses with thousands of taxa and mixed models.
Stamatakis A, Ludwig T, Meier H. 2005. RAxML-III: a fast program for
maximum likelihood-based inference of large phylogenetic trees.
Swofford DL. 1998. PAUP*: phylogenetic analysis using parsimony
(and other methods). Sunderland, MA: Sinauer Associates.
Tajima F. 1993. Simple methods for testing the molecular
evolutionary clock hypothesis. Genetics 135:599–607.
Takezaki N, Rzhetsky A, Nei M. 1995. Phylogenetic test of
Tamura K, Nei M. 1993. Estimation of the number of nucleotide
substitutions in the control region of mitochondrial DNA in
humans and chimpanzees. Mol Biol Evol. 10:512–526.
Tamura K, Nei M, Kumar S. 2004. Prospects for inferring very large
phylogenies by using the neighbor-joining method. Proc Natl
Acad Sci U S A. 101:11030–11035.
Yang Z. 1994. Maximum likelihood phylogenetic estimation from
DNA sequences with variable rates over sites: approximate
methods. J Mol Evol. 39:306–314.
Yang Z. 2006. Computational molecular evolution. Oxford: Oxford
Yang Z, Kumar S, Nei M. 1995. A new method of inference of
ancestral nucleotide and amino acid sequences. Genetics
MEGA5: Molecular Evolutionary Genetics Analysis · doi:10.1093/molbev/msr121