# Tutorial

This is the manual for version 3.1. Please use the new version.

# Tutorial: A Simple Analysis

This section is a tutorial based on the primates.nex data file. It will guide you through a basic Bayesian MCMC analysis of phylogeny, explaining the most important features of the program. There are two versions of the tutorial. You will first find a Quick-Start version for impatient users who want to get an analysis started immediately. The rest of the section contains a much more detailed description of the same analysis.

## Quick Start Version

There are four steps to a typical Bayesian phylogenetic analysis using MrBayes:

1. Read the Nexus data file
2. Set the evolutionary model
3. Run the analysis
4. Summarize the samples

In more detail, each of these steps is performed as described in the following paragraphs.

1. At the MrBayes > prompt, type execute primates.nex. This will bring the data into the program. The data file (primates.nex) must be in the same directory as the MrBayes program. If you run into problems, refer to section 2.2 for possible solutions. If you are running your own data file, beware that it may contain some MrBayes commands that can change the behavior of the program; delete those commands or put them in square brackets to follow this tutorial.

2. At the MrBayes > prompt, type lset nst=6 rates=invgamma. This sets the evolutionary model to the GTR model with gamma-distributed rate variation across sites and a proportion of invariable sites. If your data are not DNA or RNA, if you want to invoke a different model, or if you want to use non-default priors, refer to the rest of this manual, particularly sections 3 to 5 and the Appendix.

3.1. At the MrBayes > prompt, type mcmc ngen=10000 samplefreq=10. This will ensure that you get at least 1,000 samples from the posterior probability distribution. For larger data sets you probably want to run the analysis longer and sample less frequently (the default sampling frequency is every 100th generation). You can find the predicted remaining time to completion of the analysis in the last column printed to screen.

3.2. If the standard deviation of split frequencies is below 0.01 after 100,000 generations, stop the run by answering no when the program asks "Continue the analysis? (yes/no)". Otherwise, keep adding generations until the value falls below 0.01.

4.1. Summarize the parameter values by typing sump burnin=250 (or whatever value corresponds to 25 % of your samples). The program will output a table with summaries of the samples of the substitution model parameters, including the mean, mode, and 95 % credibility interval of each parameter. Make sure that the potential scale reduction factor (PSRF) is reasonably close to 1.0 for all parameters; if not, you need to run the analysis longer.

4.2. Summarize the trees by typing sumt burnin=250 (or whatever value corresponds to 25 % of your samples). The program will output a cladogram with the posterior probabilities for each split and a phylogram with mean branch lengths. The trees will also be printed to a file that can be read by tree drawing programs such as TreeView, MacClade, and Mesquite.

It does not have to be more complicated than this; however, as you get more proficient you will probably want to know more about what is happening behind the scenes. The rest of this section explains each of the steps in more detail and introduces you to all the implicit assumptions you are making and the machinery that MrBayes uses in order to perform your analysis.

## Getting Data into MrBayes

To get data into MrBayes, you need a so-called Nexus file that contains aligned nucleotide or amino acid sequences, morphological ("standard") data, restriction site (binary) data, or any mix of these four data types. The Nexus file that we will use for this tutorial, primates.nex, contains 12 mitochondrial DNA sequences of primates. A Nexus file is a simple text (ASCII) file that begins with #NEXUS on the first line. The rest of the file is divided into different blocks. The primates.nex file looks like this:

#NEXUS
begin data;
dimensions ntax=12 nchar=898;
format datatype=dna interleave=no gap=-;
matrix
Saimiri_sciureus AAGCTTCATAGGAGC ... ACTATCCCTAAGCTT
Tarsius_syrichta AAGCTTCACCGGCGC ... ATTATGCCTAAGCTT
Lemur_catta      AAGCTTCACCGGCGC ... ACTATCTATTAGCTT
Macaca_fuscata   AAGCTTCACCGGCGC ... CCTAACGCTAAGCTT
M_mulatta        AAGCTTCACCGGCGC ... CCTAACACTAAGCTT
M_fascicularis   AAGCTTTACAGGTGC ... CCTAACACTAAGCTT
M_sylvanus       AAGCTTTTCCGGCGC ... CCTAACATTAAGCTT
Homo_sapiens     AAGCTTTTCTGGCGC ... GCTCTCCCTAAGCTT
Pan              AAGCTTCTCCGGCGC ... GCTCTCCCTAAGCTT
Gorilla          AAGCTTCTCCGGTGC ... ACTCTCCCTAAGCTT
Pongo            AAGCTTCACCGGCGC ... ACTCTCACTAAGCTT
Hylobates        AAGTTTCATTGGAGC ... ACTCTCCCTAAGCTT
;
end;


The file contains only one block, a DATA block. The DATA block is initialized with begin data; followed by the dimensions statement, the format statement, and the matrix statement. The dimensions statement must contain ntax, the number of taxa, and nchar, the number of characters in each aligned sequence. The format statement must specify the type of data, for instance datatype=DNA (or RNA or Protein or Standard or Restriction). The format statement may also contain gap=- (or whatever symbol is used for a gap in your alignment), missing=? (or whatever symbol is used for missing data in your file), interleave=yes when the data matrix is interleaved sequences, and match=. (or whatever symbol is used for matching characters in the alignment). The format statement is followed by the matrix statement, usually started by the word matrix on a separate line, followed by the aligned sequences. Each sequence begins with the sequence name separated from the sequence itself by at least one space. The data block is completed by an end; statement. Note that the begin, dimensions, format, and end statements all end in a semicolon. That semicolon is essential and must not be left out. Note that, although it occupies many lines in the file, the matrix description is also a regular statement, a matrix statement, which ends with a semicolon just as any other statement. Our example file contains sequences in non-interleaved (block) format. Non-interleaved is the default but you can put interleave=no in the format statement if you want to be sure.

The Nexus data file can be generated by a program such as MacClade or Mesquite. Note, however, that MrBayes does not support the full Nexus standard, so you may have to do a little editing of the file for MrBayes to process it properly. In particular, MrBayes uses a fixed set of symbols for each data type and does not support the options of the upported symbols are {A, C, G, T, R, Y, M, K, S, W, H, B, V, D, N} for DNA data, {A, C, G, U, R, Y, M, K, S, W, H, B, V, D, N} for RNA data, {A, R, N, D, C, Q, E, G, H, I, L, K, M, F, P, S, T, W, Y, V, X} for protein data, {0, 1} for restriction (binary) data, and {0, 1, 2, 3, 4, 5, 6, 5, 7, 8, 9} for standard (morphology) data. In addition to the standard one-letter ambiguity symbols for DNA and RNA listed above, ambiguity can also be expressed using the Nexus parenthesis or curly braces notation. For instance, a taxon polymorphic for states 2 and 3 can be coded as (23), (2,3), {23}, or {2,3} and a taxon with either amino acid A or F can be coded as (AF), (A,F), {AF} or {A,F}. Like most other statistical phylogenetics programs, MrBayes effectively treats polymorphism and uncertainty the same way (as uncertainty), so it does not matter whether you use parentheses or curly brackets. If you have other symbols in your matrix than the ones supported by MrBayes, you need to replace them before processing the data block in MrBayes. You also need to remove the "Equate" and "Symbols" statements in the "Format" line if they are included. Unlike the Nexus standard, MrBayes supports data blocks that contain mixed data types as described in section 3 of this manual.

To put the data into MrBayes type execute <filename> at the MrBayes > prompt, where <filename> is the name of the input file. To process our example file, type execute primates.nex or simply exe primates.nex to save some typing. Note that the input file must be located in the same folder (directory) as the MrBayes application (or else you will have to give the path to the file) and the name of the input file should not have blank spaces. If everything proceeds normally, MrBayes will acknowledge that it has read the data in the DATA block of the Nexus file by outputting the following information:

  Executing file "primates.nex"
DOS line termination
Longest line length = 915
Parsing file
Expecting NEXUS formatted file
Allocated matrix
Matrix has 12 taxa and 898 characters
Data is Dna
Data matrix is not interleaved
Gaps coded as -
Setting default partition (does not divide up characters).
Taxon  1 -> Tarsius_syrichta
Taxon  2 -> Lemur_catta
Taxon  3 -> Homo_sapiens
Taxon  4 -> Pan
Taxon  5 -> Gorilla
Taxon  6 -> Pongo
Taxon  7 -> Hylobates
Taxon  8 -> Macaca_fuscata
Taxon  9 -> M_mulatta
Taxon 10 -> M_fascicularis
Taxon 11 -> M_sylvanus
Taxon 12 -> Saimiri_sciureus
Setting output file names to "primates.nex.<run<i>.p/run<i>.t>"
Exiting data block
Reached end of file


## Specifying a Model

All of the commands are entered at the MrBayes > prompt. At a minimum two commands, lset and prset, are required to specify the evolutionary model that will be used in the analysis. Usually, it is also a good idea to check the model settings prior to the analysis using the showmodel command. In general, lset is used to define the structure of the model and prset is used to define the prior probability distributions on the parameters of the model. In the following, we will specify a GTR + I + _ model (a General Time Reversible model with a proportion of invariable sites and a gamma-shaped distribution of rates across sites) for the evolution of the mitochondrial sequences and we will check all of the relevant priors. If you are unfamiliar with stochastic models of molecular evolution, we suggest that you consult a general text, such as Felsenstein (2004).

In general, a good start is to type help lset. Ignore the help information for now and concentrate on the table at the bottom of the output, which specifies the current settings. It should look like this:

  Model settings for partition 1:

Parameter    Options                               Current Setting
------------------------------------------------------------------
Nucmodel     4by4/Doublet/Codon                    4by4
Nst          1/2/6                                 1
Code         Universal/Vertmt/Mycoplasma/
Yeast/Ciliates/Metmt                  Universal
Ploidy       Haploid/Diploid                       Diploid
Ngammacat    <number>                              4
Nbetacat     <number>                              5
Omegavar     Equal/Ny98/M3                         Equal
Covarion     No/Yes                                No
Coding       All/Variable/Noabsencesites/
Nopresencesites                       All
Parsmodel    No/Yes                                No
------------------------------------------------------------------


First, note that the table is headed by Model settings for partition 1. By default, MrBayes divides the data into one partition for each type of data you have in your DATA block. If you have only one type of data, all data will be in a single partition by default. How to change the partitioning of the data will be explained in section 3 of the manual.

The Nucmodel setting allows you to specify the general type of DNA model. The Doublet option is for the analysis of paired stem regions of ribosomal DNA and the Codon option is for analyzing the DNA sequence in terms of its codons. We will analyze the data using a standard nucleotide substitution model, in which case the default 4by4 option is appropriate, so we will leave Nucmodel at its default setting.

The general structure of the substitution model is determined by the Nst setting. By default, all substitutions have the same rate (Nst=1), corresponding to the F81 model (or the JC model if the stationary state frequencies are forced to be equal using the prset command, see below). We want the GTR model (Nst=6 instead of the F81 model so we type lset nst=6. MrBayes should acknowledge that it has changed the model settings.

The Code setting is only relevant if the Nucmodel is set to Codon. The Ploidy setting is also irrelevant for us. However, we need to change the Rates setting from the default Equal (no rate variation across sites) to Invgamma (gamma-shaped rate variation with a proportion of invariable sites). Do this by typing lset rates=invgamma. Again, MrBayes will acknowledge that it has changed the settings. We could have changed both lset settings at once if we had typed lset nst=6 rates=invgamma in a single line.

We will leave the Ngammacat setting (the number of discrete categories used to approximate the gamma distribution) at the default of 4. In most cases, four rate categories are sufficient. It is possible to increase the accuracy of the likelihood calculations by increasing the number of rate categories. However, the time it will take to complete the analysis will increase in direct proportion to the number of rate categories you use, and the effects on the results will be negligible in most cases.

Of the remaining settings, it is only Covarion and Parsmodel that are relevant for single nucleotide models. We will use neither the parsimony model nor the covariotide model for our data, so we will leave these settings at their default values. If you type help lset now to verify that the model is correctly set, the table should look like this:

  Model settings for partition 1:

Parameter    Options                               Current Setting
------------------------------------------------------------------
Nucmodel     4by4/Doublet/Codon                    4by4
Nst          1/2/6                                 6
Code         Universal/Vertmt/Mycoplasma/
Yeast/Ciliates/Metmt                  Universal
Ploidy       Haploid/Diploid                       Diploid
Ngammacat    <number>                              4
Nbetacat     <number>                              5
Omegavar     Equal/Ny98/M3                         Equal
Covarion     No/Yes                                No
Coding       All/Variable/Noabsencesites/
Nopresencesites                       All
Parsmodel    No/Yes                                No
------------------------------------------------------------------


## Setting the Priors

We now need to set the priors for our model. There are six types of parameters in the model: the topology, the branch lengths, the four stationary frequencies of the nucleotides, the six different nucleotide substitution rates, the proportion of invariable sites, and the shape parameter of the gamma distribution of rate variation. The default priors in MrBayes work well for most analyses, and we will not change any of them here so if you are impatient, you can skip this step and continue with Start reviewing this information by typing help prset to obtain a list of the default settings for the parameters in your model. The table at the end of the help information reads:

  Model settings for partition 1:

Parameter        Options                      Current Setting
------------------------------------------------------------------
Tratiopr         Beta/Fixed                   Beta(1.0,1.0)
Revmatpr         Dirichlet/Fixed              Dirichlet(1.0,1.0,1.0,1.0,1.0,1.0)
Aamodelpr        Fixed/Mixed                  Fixed(Poisson)
Aarevmatpr       Dirichlet/Fixed              Dirichlet(1.0,1.0,...)
Omegapr          Dirichlet/Fixed              Dirichlet(1.0,1.0)
Ny98omega1pr     Beta/Fixed                   Beta(1.0,1.0)
Ny98omega3pr     Uniform/Exponential/Fixed    Exponential(1.0)
M3omegapr        Exponential/Fixed            Exponential
Codoncatfreqs    Dirichlet/Fixed              Dirichlet(1.0,1.0,1.0)
Statefreqpr      Dirichlet/Fixed              Dirichlet(1.0,1.0,1.0,1.0)
Ratepr           Fixed/Variable=Dirichlet     Fixed
Shapepr          Uniform/Exponential/Fixed    Uniform(0.0,50.0)
Ratecorrpr       Uniform/Fixed                Uniform(-1.0,1.0)
Pinvarpr         Uniform/Fixed                Uniform(0.0,1.0)
Covswitchpr      Uniform/Exponential/Fixed    Uniform(0.0,100.0)
Symmetricbetapr  Uniform/Exponential/Fixed    Fixed(Infinity)
Topologypr       Uniform/Constraints          Uniform
Brlenspr         Unconstrained/Clock          Unconstrained:Exp(10.0)
Speciationpr     Uniform/Exponential/Fixed    Uniform(0.0,10.0)
Extinctionpr     Uniform/Exponential/Fixed    Uniform(0.0,10.0)
Sampleprob       <number>                     1.00
Thetapr          Uniform/Exponential/Fixed    Uniform(0.0,10.0)
Growthpr         Uniform/Exponential/
Fixed/Normal                 Fixed(0.0)
------------------------------------------------------------------


We need to focus on Revmatpr (for the six substitution rates of the GTR rate matrix), Statefreqpr (for the stationary nucleotide frequencies of the GTR rate matrix), Shapepr (for the shape parameter of the gamma distribution of rate variation), Pinvarpr (for the proportion of invariable sites), Topologypr (for the topology), and Brlenspr (for the branch lengths).

The default prior probability density is a flat Dirichlet (all values are 1.0) for both Revmatpr and Statefreqpr. This is appropriate if we want estimate these parameters from the data assuming no prior knowledge about their values. It is possible to fix the rates and nucleotide frequencies but this is generally not recommended. However, it is occasionally necessary to fix the nucleotide frequencies to be equal, for instance in specifying the JC and SYM models. This would be achieved by typing prset statefreqpr=fixed(equal).

If we wanted to specify a prior that put more emphasis on equal nucleotide frequencies than the default flat Dirichlet prior, we could for instance use prset statefreqpr = Dirichlet(10,10,10,10) or, for even more emphasis on equal frequencies, prset statefreqpr=Dirichlet(100,100,100,100). The sum of the numbers in the Dirichlet distribution determines how focused the distribution is, and the balance between the numbers determines the expected proportion of each nucleotide (in the order A, C, G, and T). Usually, there is a connection between the parameters in the Dirichlet distribution and observations. Thus, you can think of a Dirichlet (150,100,90,140) distribution as one arising from observing 150 100 Ce set of sequences. If that set of sequences were independent of your set of sequences but clearly relevant to the analysis of your sequences, it might be reasonable to use those numbers as a prior in your analysis.

In our analysis, we will be cautious and leave the prior on state frequencies at its default setting. If you have changed the setting according to the suggestions above, you need to change it back by typing prset statefreqpr=Dirichlet(1,1,1,1) or prs st = Dir(1,1,1,1) if you want to save some typing. Similarly, we will leave the prior on the substitution rates at the default flat Dirichlet(1,1,1,1,1,1) distribution.

The Shapepr parameter determines the prior for the $\alpha$(shape) parameter of the gamma distribution of rate variation. We will leave it at its default setting, a uniform distribution spanning a wide range of $\alpha$values. The prior for the proportion of invariable sites is set with Pinvarpr. The default setting is a uniform distribution between 0 and 1, an appropriate setting if we don't want to assume any prior knowledge about the proportion of invariable sites.

For topology, the default Uniform setting for the Topologypr parameter puts equal probability on all distinct, fully resolved topologies. The alternative is to constrain some nodes in the tree to always be present but we will not attempt that in this analysis.

The Brlenspr parameter can either be set to unconstrained or clock-constrained. For trees without a molecular clock (unconstrained) the branch length prior can be set either to exponential or uniform. The default exponential prior with parameter 10.0 should work well for most analyses. It has an expectation of 1/10 = 0.1 but allows a wide range of branch length values (theoretically from 0 to infinity). Because the likelihood values vary much more rapidly for short branches than for long branches, an exponential prior on branch lengths is closer to being uninformative than a uniform prior.

## Checking the Model

To check the model before we start the analysis, type showmodel. This will give an overview of the model settings. In our case, the output will be as follows:

  Model settings:

Datatype  = DNA
Nucmodel  = 4by4
Nst       = 6
Substitution rates, expressed as proportions
of the rate sum, have a Dirichlet prior
(1.00,1.00,1.00,1.00,1.00,1.00)
Covarion  = No
# States  = 4
State frequencies have a Dirichlet prior
(1.00,1.00,1.00,1.00)
Rates     = Invgamma
Gamma shape parameter is uniformly dist-
ributed on the interval (0.05,50.00).
Proportion of invariable sites is uniformly dist-
ributed on the interval (0.00,1.00).
Gamma distribution is approximated using 4 categories.

Active parameters:

Parameters
------------------
Revmat           1
Statefreq        2
Shape            3
Pinvar           4
Topology         5
Brlens           6
------------------

1 --  Parameter  = Revmat
Prior      = Dirichlet(1.00,1.00,1.00,1.00,1.00,1.00)
2 --  Parameter  = Statefreq
Prior      = Dirichlet
3 --  Parameter  = Shape
Prior      = Uniform(0.05,50.00)
4 --  Parameter  = Pinvar
Prior      = Uniform(0.00,1.00)
5 --  Parameter  = Topology
Prior      = All topologies equally probable a priori
6 --  Parameter  = Brlens
Prior      = Branch lengths are Unconstrained:Exponential(10.0)


Note that we have six types of parameters in our model. All of these parameters will be estimated during the analysis (see section 5 for information on how to fix parameter values).

## Setting up the Analysis

The analysis is started by issuing the mcmc command. However, before doing this, we recommend that you review the run settings by typing help mcmc. The help mcmc command will produce the following table at the bottom of the output:

  Parameter       Options               Current Setting
-----------------------------------------------------
Seed            <number>              1115403472
Swapseed        <number>              1115403472
Ngen            <number>              1000000
Nruns           <number>              2
Nchains         <number>              4
Temp            <number>              0.200000
Reweight        <number>,<number>     0.00 v 0.00 ^
Swapfreq        <number>              1
Nswaps          <number>              1
Samplefreq      <number>              100
Printfreq       <number>              100
Printall        Yes/No                Yes
Printmax        <number>              8
Mcmcdiagn       Yes/No                Yes
Diagnfreq       <number>              1000
Minpartfreq     <number>              0.10
Allchains       Yes/No                No
Allcomps        Yes/No                No
Relburnin       Yes/No                Yes
Burnin          <number>              0
Burninfrac      <number>              0.25
Stoprule        Yes/No                No
Stopval         <number>              0.01
Filename        <name>                primates.nex.<p/t>
Startingtree    Random/User           Random
Nperts          <number>              0
Savebrlens      Yes/No                Yes
Ordertaxa       Yes/No                No
---------------------------------------------------------------------------


The Seed is simply the seed for the random number generator, and Swapseed is the seed for the separate random number generator used to generate the chain swapping sequence (see below). Unless they are set to user-specified values, these seeds are generated from the system clock, so your values are likely to be different from the ones in the screen dump above. The Ngen setting is the number of generations for which the analysis will be run. It is useful to run a small number of generations first to make sure the analysis is correctly set up and to get an idea of how long it will take to complete a longer analysis. We will start with 10,000 generations. To change the Ngen setting without starting the analysis we use the mcmcp command, which is equivalent to mcmc except that it does not start the analysis. Type mcmcp ngen=10000 to set the number of generations to 10,000. You can type help mcmc to confirm that the setting was changed appropriately.

By default, MrBayes will run two simultaneous, completely independent analyses starting from different random trees (Nruns = 2). Running more than one analysis simultaneously is very helpful in determining when you have a good sample from the posterior probability distribution, so we suggest that you leave this setting as is. The idea is to start each run from different randomly chosen trees. In the early phases of the run, the two runs will sample very different trees but when they have reached convergence (when they produce a good sample from the posterior probability distribution), the two tree samples should be very similar.

To make sure that MrBayes compares tree samples from the different runs, make sure that Mcmcdiagn is set to yes and that Diagnfreq is set to some reasonable value, such as every 1000th generation. MrBayes will now calculate various run diagnostics every Diagnfreq generation and print them to a file with the name <Filename>.mcmc. The most important diagnostic, a measure of the similarity of the tree samples in the different runs, will also be printed to screen every Diagnfreq generation. Every time the diagnostics are calculated, either a fixed number of samples (burnin) or a percentage of samples (burninfrac) from the beginning of the chain is discarded. The relburnin setting determines whether a fixed burnin (relburnin=no) or a burnin percentage (relburnin=yes) is used. If you , MrBayes will by default discard the first 25 % samples from the cold chain (relburnin=yes and burninfrac=0.25).

By default, MrBayes uses Metropolis coupling to improve the MCMC sampling of the target distribution. The Swapfreq, Nswaps, Nchains, and Temp settings together control the Metropolis coupling behavior. When Nchains is set to 1, no heating is used. When Nchains is set to a value n larger than 1, then n - 1 heated chains are used. By default, Nchains is set to 4, meaning that MrBayes will use 3 heated chains and one "cold" chain. In our experience, heating is essential for problems with more than about 50 taxa, whereas smaller problems often can be analyzed successfully without heating. Adding more than three heated chains may be helpful in analyzing large and difficult data sets. The time complexity of the analysis is directly proportional to the number of chains used (unless MrBayes runs out of physical RAM memory, in which case the analysis will suddenly become much slower), but the cold and heated chains can be distributed among processors in a cluster of computers using the MPI version of the program (see below), greatly speeding up the calculations.

MrBayes uses an incremental heating scheme, in which chain $i$ is heated by raising its posterior probability to the power $1/ (1 + i\lambda$), where $\lambda$ is the temperature controlled by the Temp parameter. The effect of the heating is to flatten out the posterior probability, such that the heated chains more easily find isolated peaks in the posterior distribution and can help the cold chain move more rapidly between these peaks. Every Swapfreq generation, two chains are picked at random and an attempt is made to swap their states. For many analyses, the default settings should work nicely. If you are running many more than three heated chains, however, you may want to increase the number of swaps (Nswaps) that are tried each time the chain stops for swapping.

The Samplefreq setting determines how often the chain is sampled. By default, the chain is sampled every 100th generation, and this works well for most analyses. However, our analysis is so small that we are likely to get convergence quickly, so it makes sense to sample the chain more frequently, say every 10th generation (this will ensure that we get at least 1,000 samples when the number of generations is set to 10,000). To change the sampling frequency, type mcmcp samplefreq=10.

When the chain is sampled, the current values of the model parameters are printed to file. The substitution model parameters are printed to a .p file (in our case, there will be one file for each independent analysis, and they will be called primates.nex.run1.p and primates.nex.run2.p). The .p files are tab delimited text files that can be imported into most statistics and graphing programs. The topology and branch lengths are printed to a .t file (in our case, there will be two files called primates.nex.run1.t and primates.nex.run2.t). The .t files are Nexus tree files that can be imported into programs like PAUP* and TreeView. The root of the .p and .t file names can be altered using the Filename setting.

The Printfreq parameter controls the frequency with which the state of the chains is printed to screen. Leave Printfreq at the default value (print to screen every 100th generation).

The default behavior of MrBayes is to save trees with branch lengths to the .t file. Since this is what we want, we leave this setting as is. If you are running a large analysis (many taxa) and are not interested in branch lengths, you can save a considerable amount of disk space by not saving branch lengths.

The Startingtree parameter can be used to feed the chain(s) with a user-specified starting tree. The default behavior is to start each chain with a different random tree; this is recommended for general use.

## Running the Analysis

Finally, we are ready to start the analysis. Type mcmc. MrBayes will first print information about the model and then list the proposal mechanisms that will be used in sampling from the posterior distribution. In our case, the proposals are the following:

  The MCMC sampler will use the following moves:
With prob.  Chain will change
4.17 %   param. 1 (revmat) with Dirichlet proposal
4.17 %   param. 2 (state frequencies) with Dirichlet proposal
4.17 %   param. 3 (gamma shape) with multiplier
4.17 %   param. 4 (prop. invariants) with beta proposal
20.83 %   param. 5 (topology and branch lengths) with LOCAL
62.50 %   param. 5 (topology and branch lengths) with extending TBR


Note that MrBayes will spend most of its effort changing topology and branch lengths. In our experience, topology and branch lengths are the most difficult parameters to integrate over and we therefore let MrBayes spend a large proportion of its time proposing new values for those parameters. The proposal probabilities can be changed with the props command but be warned that inappropriate changes of proposal probabilities may destroy any hopes of achieving convergence. After the initial log likelihoods, MrBayes will print the state of the chains every 100th generation, like this:

  Chain results:

1 -- (-7812.831) (-7523.685) [-7485.569] (-7700.309) * [-7832.045] (-7618.595) (-7776.608) (-7836.826)
100 -- (-6771.532) (-6857.529) (-6766.678) [-6682.527] * [-6506.277] (-6944.449) (-6784.126) (-6991.307) -- 0:01:39
200 -- (-6321.464) [-6179.561] (-6338.168) (-6272.242) * (-6339.400) (-6715.840) [-6265.329] (-6599.698) -- 0:01:38
300 -- (-6201.285) (-6084.899) (-6139.200) [-6049.061] * [-6073.056] (-6359.134) (-6106.834) (-6515.348) -- 0:01:37
400 -- (-6015.429) [-5924.177] (-6074.274) (-5967.236) * (-6002.941) (-6215.415) [-5980.096] (-6036.624) -- 0:01:12
500 -- (-5963.069) [-5851.415] (-6019.301) (-5948.768) * [-5925.423] (-6072.446) (-5925.549) (-5963.427) -- 0:01:16
600 -- (-5931.545) [-5802.472] (-5997.276) (-5927.204) * (-5899.590) (-6035.097) (-5904.327) [-5896.520] -- 0:01:18
700 -- (-5852.405) [-5782.934] (-5966.224) (-5880.975) * (-5890.220) (-6008.142) (-5876.956) [-5838.392] -- 0:01:19
800 -- (-5844.426) [-5754.416] (-5963.531) (-5860.182) * (-5883.474) (-5902.784) (-5790.765) [-5789.326] -- 0:01:20
900 -- (-5814.885) [-5741.101] (-5926.666) (-5855.916) * (-5870.806) (-5871.573) (-5783.012) [-5755.608] -- 0:01:20
1000 -- (-5811.949) [-5737.884] (-5888.234) (-5819.793) * (-5867.377) (-5851.693) (-5784.437) [-5749.264] -- 0:01:21

Average standard deviation of split frequencies: 0.073946

1100 -- (-5784.586) [-5730.069] (-5880.476) (-5798.728) * (-5839.268) (-5836.074) (-5779.940) [-5739.170] -- 0:01:20

...

9000 -- (-5737.756) [-5728.228] (-5735.787) (-5722.263) * [-5730.926] (-5740.655) (-5730.045) (-5729.487) -- 0:00:09

Average standard deviation of split frequencies: 0.000000

9100 -- (-5732.093) (-5733.215) (-5735.816) [-5720.361] * (-5737.343) (-5741.841) (-5735.730) [-5726.010] -- 0:00:08
9200 -- (-5731.412) (-5728.251) (-5729.846) [-5720.670] * (-5731.852) (-5740.270) (-5742.079) [-5729.764] -- 0:00:07
9300 -- (-5739.416) (-5734.909) [-5723.707] (-5720.408) * (-5728.016) (-5728.550) [-5730.645] (-5726.709) -- 0:00:06
9400 -- [-5720.648] (-5732.631) (-5732.691) (-5724.562) * (-5728.739) [-5720.959] (-5739.326) (-5720.832) -- 0:00:05
9500 -- (-5730.870) (-5729.773) (-5729.010) [-5727.405] * (-5727.274) [-5720.702] (-5737.968) (-5726.367) -- 0:00:04
9600 -- (-5736.460) (-5734.016) (-5737.829) [-5727.732] * (-5728.393) [-5718.892] (-5736.313) (-5727.928) -- 0:00:03
9700 -- (-5740.378) (-5733.312) (-5735.204) [-5724.199] * (-5728.102) [-5724.006] (-5736.798) (-5722.292) -- 0:00:02
9800 -- (-5739.019) (-5735.541) (-5731.738) [-5728.466] * (-5723.404) (-5722.901) (-5734.184) [-5722.141] -- 0:00:01
9900 -- (-5743.556) [-5724.246] (-5736.484) (-5731.803) * [-5727.082] (-5732.476) (-5728.482) (-5725.045) -- 0:00:00
10000 -- (-5738.106) [-5723.380] (-5726.322) (-5727.599) * (-5728.776) (-5732.410) [-5728.515] (-5725.418) -- 0:00:00

Average standard deviation of split frequencies: 0.000000

Continue with analysis? (yes/no):


If you have the terminal window wide enough, each generation of the chain will print on a single line like this (we have decreased the font size to show the layout):

  Chain results:

1 -- (-7812.831) (-7523.685) [-7485.569] (-7700.309) * [-7832.045] (-7618.595) (-7776.608) (-7836.826)
100 -- (-6771.532) (-6857.529) (-6766.678) [-6682.527] * [-6506.277] (-6944.449) (-6784.126) (-6991.307) -- 0:01:39
200 -- (-6321.464) [-6179.561] (-6338.168) (-6272.242) * (-6339.400) (-6715.840) [-6265.329] (-6599.698) -- 0:01:38
300 -- (-6201.285) (-6084.899) (-6139.200) [-6049.061] * [-6073.056] (-6359.134) (-6106.834) (-6515.348) -- 0:01:37
400 -- (-6015.429) [-5924.177] (-6074.274) (-5967.236) * (-6002.941) (-6215.415) [-5980.096] (-6036.624) -- 0:01:12
500 -- (-5963.069) [-5851.415] (-6019.301) (-5948.768) * [-5925.423] (-6072.446) (-5925.549) (-5963.427) -- 0:01:16
600 -- (-5931.545) [-5802.472] (-5997.276) (-5927.204) * (-5899.590) (-6035.097) (-5904.327) [-5896.520] -- 0:01:18
700 -- (-5852.405) [-5782.934] (-5966.224) (-5880.975) * (-5890.220) (-6008.142) (-5876.956) [-5838.392] -- 0:01:19
800 -- (-5844.426) [-5754.416] (-5963.531) (-5860.182) * (-5883.474) (-5902.784) (-5790.765) [-5789.326] -- 0:01:20
900 -- (-5814.885) [-5741.101] (-5926.666) (-5855.916) * (-5870.806) (-5871.573) (-5783.012) [-5755.608] -- 0:01:20
1000 -- (-5811.949) [-5737.884] (-5888.234) (-5819.793) * (-5867.377) (-5851.693) (-5784.437) [-5749.264] -- 0:01:21

Average standard deviation of split frequencies: 0.073946

1100 -- (-5784.586) [-5730.069] (-5880.476) (-5798.728) * (-5839.268) (-5836.074) (-5779.940) [-5739.170] -- 0:01:20

...

9000 -- (-5737.756) [-5728.228] (-5735.787) (-5722.263) * [-5730.926] (-5740.655) (-5730.045) (-5729.487) -- 0:00:09

Average standard deviation of split frequencies: 0.000000

9100 -- (-5732.093) (-5733.215) (-5735.816) [-5720.361] * (-5737.343) (-5741.841) (-5735.730) [-5726.010] -- 0:00:08
9200 -- (-5731.412) (-5728.251) (-5729.846) [-5720.670] * (-5731.852) (-5740.270) (-5742.079) [-5729.764] -- 0:00:07
9300 -- (-5739.416) (-5734.909) [-5723.707] (-5720.408) * (-5728.016) (-5728.550) [-5730.645] (-5726.709) -- 0:00:06
9400 -- [-5720.648] (-5732.631) (-5732.691) (-5724.562) * (-5728.739) [-5720.959] (-5739.326) (-5720.832) -- 0:00:05
9500 -- (-5730.870) (-5729.773) (-5729.010) [-5727.405] * (-5727.274) [-5720.702] (-5737.968) (-5726.367) -- 0:00:04
9600 -- (-5736.460) (-5734.016) (-5737.829) [-5727.732] * (-5728.393) [-5718.892] (-5736.313) (-5727.928) -- 0:00:03
9700 -- (-5740.378) (-5733.312) (-5735.204) [-5724.199] * (-5728.102) [-5724.006] (-5736.798) (-5722.292) -- 0:00:02
9800 -- (-5739.019) (-5735.541) (-5731.738) [-5728.466] * (-5723.404) (-5722.901) (-5734.184) [-5722.141] -- 0:00:01
9900 -- (-5743.556) [-5724.246] (-5736.484) (-5731.803) * [-5727.082] (-5732.476) (-5728.482) (-5725.045) -- 0:00:00
10000 -- (-5738.106) [-5723.380] (-5726.322) (-5727.599) * (-5728.776) (-5732.410) [-5728.515] (-5725.418) -- 0:00:00

Average standard deviation of split frequencies: 0.000000

Continue with analysis? (yes/no):


The first column lists the generation number. The following four columns with negative numbers each correspond to one chain in the first run. Each column corresponds to one physical location in computer memory, and the chains actually shift positions in the columns as the run proceeds. The numbers are the log likelihood values of the chains. The chain that is currently the cold chain has its value surrounded by square brackets, whereas the heated chains have their values surrounded by parentheses. When two chains successfully change states, they trade column positions (places in computer memory). If the Metropolis coupling works well, the cold chain should move around among the columns; this means that the cold chain successfully swaps states with the heated chains. If the cold chain gets stuck in one of the columns, then the heated chains are not successfully contributing states to the cold chain, and the Metropolis coupling is inefficient. The analysis may then have to be run longer or the temperature difference between chains may have to be lowered.

The star column separates the two different runs. The last column gives the time left to completion of the specified number of generations. This analysis approximately takes 1 second per 100 generations. Because different moves are used in each generation, the exact time varies somewhat for each set of 100 generations, and the predicted time to completion will be unstable in the beginning of the run. After a while, the predictions will become more accurate and the time will decrease predictably.

## When to Stop the Analysis

At the end of the run, MrBayes asks whether or not you want to continue with the analysis. Before answering that question, examine the average standard deviation of split frequencies. As the two runs converge onto the stationary distribution, we expect the average standard deviation of split frequencies to approach zero, reflecting the fact that the two tree samples become increasingly similar. In our case, the average standard deviation is about 0.07 after 1,000 generations and then drops to less than 0.000001 towards the end of the run. Your values can differ slightly because of stochastic effects. Given the extremely low value of the average standard deviation at the end of the run, there appears to be no need to continue the analysis beyond 10,000 generations so when MrBayes asks "Continue with analysis? (yes/no):", stop the analysis by typing no.

Although we recommend using a convergence diagnostic, such as the standard deviation of split frequencies, to determine run length, we want to mention that there are simpler but less powerful methods of determining when to stop the analysis. Arguably the simplest technique is to examine the log likelihood values (or, more exactly, the log probability of the data given the parameter values) of the cold chain, that is, the values printed to screen within square brackets. In the beginning of the run, the values typically increase rapidly (the absolute values decrease, since these are negative numbers). This phase of the run is referred to as the "burn-in" and the samples from this phase are typically discarded. Once the likelihood of the cold chain stops to increase and starts to randomly fluctuate within a more or less stable range, the run may have reached stationarity, that is, it is producing a good sample from the posterior probability distribution. At stationarity, we also expect different, independent runs to sample similar likelihood values. Trends in likelihood values can be deceiving though; you're more likely to detect problems with convergence by comparing split frequencies than by looking at likelihood trends.

When you stop the analysis, MrBayes will print several types of information useful in optimizing the analysis. This is primarily of interest if you have difficulties in obtaining convergence. Since we apparently have a good sample from the posterior distribution already, we will ignore this information for now. We will return to the subject of optimizing the MCMC analysis in section 5 of the manual.

## Summarizing Samples of Substitution Model Parameters

During the run, samples of the substitution model parameters have been written to the .p files every samplefreq generation. These files are tab-delimited text files that look something like this:

[ID: 5848203808]
Gen  LnL        TL     r(A<->C)  ...  pi(G)     pi(T)     alpha     pinvar
1  -7559.137  2.044  0.166667  ...  0.250000  0.250000  0.500000  0.000000
10  -6585.519  2.181  0.090180  ...  0.169513  0.247418  0.569271  0.036162
...
9990  -5728.935  2.527  0.051106  ...  0.076213  0.262107  0.932771  0.194004
10000  -5722.857  2.642  0.051106  ...  0.078383  0.246791  0.721217  0.185486


The first number, in square brackets, is a randomly generated ID number that lets you identify the analysis from which the samples come. The next line contains the column headers, and is followed by the sampled values. From left to right, the columns contain: (1) the generation number (Gen); (2) the log likelihood of the cold chain (LnL); (3) the total tree length (the sum of all branch lengths, TL); (4) the six GTR rate parameters (r(A<->C), r(A<->G) etc); (5) the four stationary nucleotide frequencies (pi(A), pi(C) etc); (6) the shape parameter of the gamma distribution of rate variation (alpha); and (7) the proportion of invariable sites (pinvar). If you use a different model for your data set, the .p files will of course be different.

MrBayes provides the sump command to summarize the sampled parameter values. Before using it, we need to decide on the burn-in. Since the convergence diagnostic we used previously to determine when to stop the analysis discarded the first 25 % of the samples and indicated that convergence had been reached after 10,000 generations, it makes sense to discard 25 % of the samples obtained during the first 10,000 generations. Since we sampled every 10th generation, there are 1,000 samples (1,001 to be exact, since the first generation is always sampled) and 25 % translates to 250 samples. Thus, summarize the information in the .p file by typing sump burnin=250. By default, sump will summarize the information in the .p file generated most recently, but the filename can be changed if necessary.

The sump command will first generate a plot of the generation versus the log probability of the data (the log likelihood values). If we are at stationarity, this plot should look like "white noise", that is, there should be no tendency of increase or decrease over time. The plot should look something like this:

  +------------------------------------------------------------+ -5716.56
|     1                        1                             |
|       2             1      12               1           1 2|
|    1             2          1   1     1                 2  |
|   1           22         2    1     1    2               2 |
| 2    2   2  21        2        1        1  2               |
|   22               2 1   1             1  1   12    1      |
|2     1 **1   2  2              2 * 2 2       1211        1 |
| 11        1       *     *     2      1          2          |
|  2  2 1       1  1 12             112     2       1    2  1|
|1           2           1  *2 2  2      2 1   2   2 22  1   |
|            1   1       2          2     2        1    2    |
|                 1                     2     2      1       |
|                      21                              11    |
|           2 1                                        2     |
|                                            1      2        |
+------+-----+-----+-----+-----+-----+-----+-----+-----+-----+ -5733.79
^                                                            ^
2500                                                         10000


If you see an obvious trend in your plot, either increasing or decreasing, you probably need to run the analysis longer to get an adequate sample from the posterior probability distribution.

At the bottom of the sump output, there is a table summarizing the samples of the parameter values:

  Model parameter summaries over the runs sampled in files
"primates.nex.run1.p" and "primates.nex.run2.p":
(Summaries are based on a total of 1502 samples from 2 runs)
(Each run produced 1001 samples of which 751 samples were included)

                                             95% Cred. Interval
----------------------
Parameter       Mean        Variance       Lower         Upper         Median       PSRF *
-----------------------------------------------------------------------------------------
TL            2.885314      0.071446      2.408000      3.464000      2.861000      1.004
r(A<->C)      0.045110      0.000070      0.031406      0.061424      0.044756      0.999
r(A<->G)      0.477554      0.002421      0.387141      0.569168      0.473521      1.040
r(A<->T)      0.038541      0.000060      0.024204      0.053787      0.038402      0.999
r(C<->G)      0.033765      0.000195      0.010836      0.064049      0.032475      1.033
r(C<->T)      0.386144      0.001735      0.311504      0.459011      0.386897      1.012
r(G<->T)      0.018885      0.000133      0.001217      0.045924      0.016801      1.048
pi(A)         0.355301      0.000162      0.329769      0.377153      0.356808      1.092
pi(C)         0.320831      0.000131      0.299953      0.341757      0.321312      1.152
pi(G)         0.081154      0.000048      0.068816      0.098442      0.081723      1.014
pi(T)         0.242714      0.000104      0.226573      0.264684      0.242113      1.000
alpha         0.714405      0.037305      0.419210      1.147286      0.673545      1.007
pinvar        0.185901      0.004887      0.025889      0.307468      0.188027      1.009
-----------------------------------------------------------------------------------------
* Convergence diagnostic (PSRF = Potential scale reduction factor [Gelman
and Rubin, 1992], uncorrected) should approach 1 as runs converge. The
values may be unreliable if you have a small number of samples. PSRF should
only be used as a rough guide to convergence since all the assumptions
that allow one to interpret it as a scale reduction factor are not met in
the phylogenetic context.


For each parameter, the table lists the mean and variance of the sampled values, the lower and upper boundaries of the 95 % credibility interval, and the median of the sampled values. The parameters are the same as those listed in the .p files: the total tree length (TL), the six reversible substitution rates (r(A<->C), r(A<->G), etc), the four stationary state frequencies (pi(A), pi(C), etc), the shape of the gamma distribution of rate variation across sites (alpha), and the proportion of invariable sites (pinvar). Note that the six rate parameters of the GTR model are given as proportions of the rate sum (the Dirichlet parameterization). This parameterization has some advantages in the Bayesian context; in particular, it allows convenient formulation of priors. If you want to scale the rates relative to the G-T rate, just divide all rate proportions by the G-T rate proportion.

The last column in the table contains a convergence diagnostic, the Potential Scale Reduction Factor (PSRF). If we have a good sample from the posterior probability distribution, these values should be close to 1.0. If you have a small number of samples, there may be some spread in these values, indicating that you may need to sample the analysis more often or run it longer. In our case, we can probably easily obtain more accurate estimates of some parameters by running the analysis slightly longer.

## Summarizing Samples of Trees and Branch Lengths

Trees and branch lengths are printed to the .t files. These files are Nexus-formatted tree files with a structure like this:

#NEXUS
[ID: 5848203808]
begin trees;
translate
1 Lemur_catta,
2 Homo_sapiens,
3 Pan,
4 Gorilla,
5 Pongo,
6 Hylobates,
7 Macaca_fuscata,
8 M_mulatta,
9 M_fascicularis,
10 M_sylvanus,
11 Saimiri_sciureus,
12 Tarsius_syrichta;
tree rep.1 = ((((((2:0.100000,(8:0.100000,7:0.000748):0.197415):0.100000,(5:0.100000,6:0.100000):0.100000):0.100000,(9:0.100000,(10:0.100000,3:0.100000):0.100000):0.100000):0.100000,12:0.100000):0.100000,11:0.100000):0.100000,4:0.100000,1:0.100000);


...

  tree rep.10000 = ((((6:0.126475,((4:0.079300,(3:0.049876,2:0.066929):0.040517):0.103191,5:0.138220):0.105224):0.147212,(10:0.059671,((8:0.035939,7:0.012332):0.037992,9:0.069477):0.060859):0.221336):0.149243,11:0.627561):0.180115,12:0.525606,1:0.339476);
end;


To summarize the tree and branch length information, type sumt burnin=250. The sumt and sump commands each have separate burn-in settings so it is necessary to give the burn-in here again. Otherwise, many of the settings in MrBayes are persistent and need not be repeated every time a command is executed. To make sure the settings for a particular command are correct, you can always use help <command> before issuing the command.

The sumt command will output, among other things, summary statistics for the taxon bipartitions, a tree with clade credibility (posterior probability) values, and a phylogram (if branch lengths have been saved). The summary statistics (see below) describes each partition in the rmat (dots for the taxa that are on one side of the partition and stars for the taxa on the other side; for instance, the first partition (ID 1) is the terminal branch leading to taxon 5 since it has a star in the 5th position and a dot in all other positions). Then it gives the number of times the partition was sampled (#obs), the probability of the partition (Probab.), the standard deviation of the partition frequency (Stdev(s)), the mean (Mean(v)) and variance (Var(v)) of the branch length, the Potential Scale Reduction Factor (PSRF), and finally the number of runs in which the partition was sampled (Nruns). In our analysis, there is overwhelming support for a single tree, so all partitions have a posterior probability of 1.0.

  Summary statistics for taxon bipartitions:

ID -- Partition     #obs   Probab.  Stdev(s)   Mean(v)   Var(v)    PSRF  Nruns
------------------------------------------------------------------------------
1 -- ....*.......  1502  1.000000  0.000000  0.058828  0.000235  1.002    2
2 -- ..***.......  1502  1.000000  0.000000  0.082206  0.000426  1.000    2
3 -- ..**........  1502  1.000000  0.000000  0.030594  0.000152  0.999    2
4 -- ..*.........  1502  1.000000  0.000000  0.049234  0.000104  1.002    2
5 -- ...*........  1502  1.000000  0.000000  0.063670  0.000207  1.001    2
6 -- ..**********  1502  1.000000  0.000000  0.271433  0.003850  0.999    2
7 -- .***********  1502  1.000000  0.000000  0.472390  0.007137  1.010    2
8 -- ..*********.  1502  1.000000  0.000000  0.113412  0.002208  0.999    2
9 -- ......*.....  1502  1.000000  0.000000  0.165887  0.000861  1.017    2
10 -- ..*****.....  1502  1.000000  0.000000  0.122957  0.001186  1.009    2
11 -- ........*...  1502  1.000000  0.000000  0.022199  0.000039  1.003    2
12 -- .......*....  1502  1.000000  0.000000  0.018330  0.000042  1.002    2
13 -- ..........*.  1502  1.000000  0.000000  0.070973  0.000449  1.043    2
14 -- .*..........  1502  1.000000  0.000000  0.329111  0.003588  1.008    2
15 -- ...........*  1502  1.000000  0.000000  0.429012  0.004224  1.003    2
16 -- .........*..  1502  1.000000  0.000000  0.056534  0.000137  1.009    2
17 -- .......****.  1502  1.000000  0.000000  0.248962  0.002162  1.005    2
18 -- .....*......  1502  1.000000  0.000000  0.139867  0.000708  1.001    2
19 -- .......**...  1502  1.000000  0.000000  0.035640  0.000107  1.020    2
20 -- ..****......  1501  0.999334  0.000942  0.054652  0.000435  1.022    2
21 -- .......***..  1501  0.999334  0.000942  0.049466  0.000355  1.010    2
------------------------------------------------------------------------------


The clade credibility tree (upper tree, next page) gives the probability of each partition or clade in the tree, and the phylogram (lower tree, next page) gives the branch lengths measured in expected substitutions per site.

   Clade credibility values:

/--------------------------------------------------------- Tarsius_syrichta (1)
|
|--------------------------------------------------------- Lemur_catta (2)
|
|                                                /-------- Homo_sapiens (3)
|                                        /--100--+
|                                        |       \-------- Pan (4)
|                                /--100--+
|                                |       \---------------- Gorilla (5)
|                       /---100--+
+                       |        \------------------------ Pongo (6)
|               /--100--+
|               |       \--------------------------------- Hylobates (7)
|               |
|               |                                /-------- Macaca_fuscata (8)
|       /--100--+                        /--100--+
|       |       |                        |       \-------- M_mulatta (9)
|       |       |                /--100--+
|       |       |                |       \---------------- M_fascicularis (10)
\--100--+       \-------100------+
|                        \------------------------ M_sylvanus (11)
|
\------------------------------------------------- Saimiri_sciureus (12)

Phylogram:

/-------------------------------------- Tarsius_syrichta (1)
|
|--------------------------- Lemur_catta (2)
|
|                                                    /---- Homo_sapiens (3)
|                                                 /--+
|                                                 |  \----- Pan (4)
|                                           /-----+
|                                           |     \----- Gorilla (5)
|                                      /----+
+                                      |    \----------- Pongo (6)
|                             /--------+
|                             |        \-------------- Hylobates (7)
|                             |
|                             |                          /-- Macaca_fuscata (8)
|                    /--------+                       /--+
|                    |        |                       |  \-- M_mulatta (9)
|                    |        |                    /--+
|                    |        |                    |  \----- M_fascicularis (10)
\--------------------+        \--------------------+
|                             \----- M_sylvanus (11)
|
\---------------------------------- Saimiri_sciureus (12)

|--------| 0.100 expected changes per site


In the background, the sumt command creates three additional files. The first is a .parts file, which contains the list of taxon bipartitions, their posterior probability (the proportion of sampled trees containing them), and the branch lengths associated with them (if branch lengths have been saved). The branch length values are based only on those trees containing the relevant bipartition. The second generated file has the suffix .con and includes two consensus trees. The first one has both the posterior probability of clades (as interior node labels) and the branch lengths (if they have been saved) in its description. A graphical representation of this tree can be generated in the program TreeView by Rod Page. The second tree only contains the branch lengths and it can be imported into a wide range of tree-drawing programs such as MacClade and Mesquite. The third file generated by the sumt command is the .trprobs file, which contains the trees that were found during the MCMC search, sorted by posterior probability.