GARLI activity

table of contents

expected learning outcome

The objective of this activity is to help you learn the basics of GARLI, as well as some more advanced features. These include setting up an analysis of a nucleotide data set, monitoring a run in progress, and inspecting the output from a run. It also covers post-processing, constrained searches, bootstrap analyses, checkpointing, and codon models.

getting started

GARLI reads all of its settings from a configuration file. By default it looks for a file named garli.conf, but other configuration files can be specified as a command line argument after the executable (e.g., if the executable were named garli, you would type "garli myconfig.conf"). Note that most of the settings typically do not need to be changed from their default values, at least for exploratory runs. We will only experiment with some of the interesting settings in this demo, but detailed discussions of all settings can be found in the GARLI manual or the support web site.

The config file is divided into two parts: [general] and [master]. The [general] section specifies settings such as the data set to be read, starting conditions, model settings, and output files. To start running a new data set, the only setting that must be specified is the datafname, which specifies the file containing the data matrix in Nexus, PHYLIP, or FASTA format. The [master] section specifies settings that control the functioning of the genetic algorithm itself, and typically the default settings can be used. Basic template configuration files are included with any download of the program.

We will be using a small 29 taxon by 1218 nucleotide mammal data set that was taken from a much larger data set (44 taxa, 17kb over 20 genes) presented in (Murphy 2001). The genes included here are RAG1 and RAG2 (Recombination Activating Gene). This is a difficult data set because the internal branches are quite short, and the terminals quite long. GARLI performs less repeatably on this data set than on many others of similar size, so consider this a "worst case" data set. There are definitely local topological optima. The trees inferred using this gene also do not match our understanding of the relationships of these taxa in many places, but that is not really important for our purposes here.

Also, for this learning activity we will need to install DendroPy. Make sure you are connected to the internet and then from a terminal window, please execute the following commands:

    1. If you are running Ubuntu Linux: wget http://pypi.python.org/packages/source/D/DendroPy/DendroPy-3.9.0.tar.gz
    2. If you are running Mac OS X: curl -o DendroPy-3.9.0.tar.gz http://pypi.python.org/packages/source/D/DendroPy/DendroPy-3.9.0.tar.gz
  1. tar xvzf DendroPy-3.9.0.tar.gz
  2. cd DendroPy-3.9.0
  3. sudo python setup.py install

That's it! DendroPy is now installed. You will be using it to make consensus trees later on in the activity.

exercise 1: start a basic nucleotide run

  1. Go to the wme_jul2011/activities/GarliDemo/basicNucleotide directory
  2. Open the garli.conf file in a text editor (you can try double clicking it and if possible, tell your operating system to always open this type of file with an appropriate text editor)

There is not much that needs to be changed in the config file to start a preliminary run. In this case, a number of changes from the defaults have been made so that the example is more instructive and a bit faster (therefore, do NOT use the settings in this file as defaults for any serious analyses). You will still need to change a few things yourself. Note that the configuration file specifies that the program perform two independent search replicates (searchreps = 2). Also note that taxon 1 (Opossum) is set as the outgroup (outgroup = 1).

Make the necessary changes:

  1. Set datafname = murphy29.rag1rag2.nex
  2. Set ofprefix = run1. This will tell the program to begin the name of all output files with "run1...".
  3. Set outputphyliptree = 1. This will tell the program to write trees to file in PHYLIP as well as Nexus format.

Now we need to set the model of sequence evolution for GARLI to use. ModelTest has previously been run on this data set (you can peruse the output in the Modeltest directory). The best-fitting model under the AIC criterion is SYM+I+G (6 rates of substitution, equal base frequencies, an estimated proportion of invariant sites, gamma-distributed rate heterogeneity). This is close to the default GTR+I+G model set in the configuration file, so we only need to change the equilibrium base frequency settings.

  1. Set statefrequencies = equal
  2. Save the file
  3. Start GARLI by typing garli on the command line in the directory you want to execute the program. (It should be on your PATH, so the executable itself does not need to be in the current directory.)

You will see a bunch of text scroll by, informing you about the data set and the run that is starting. Most of this information is not very important, but if the program stops be sure to read it for any error messages. The output will also contain information about the progress of the initial optimization phase, and as the program runs it will continue to log information to the screen. This output contains the current generation number, current best lnL score, optimization precision and the generation at which the last change in topology occurred. All of the screen output is also written to a log file that in this case will be named run1.screen.log, so you can come back and look at it later.

exercise 2: monitor an ongoing run

  1. Look in the directory that GARLI is running in, and note the files that have been created by the run.
  2. Open run1.log00.log in a text editor. (You may be able to double-click it and choose to open it in an appropriate editor.)

    This logs the current best lnL, runtime, and optimization precision over the course of the run. It is useful for plotting the lnL over time. Next, we will look at the file that logs all of the information that is output to the screen.

  3. Open run1.screen.log in a text editor.

    This file contains an exact copy of the screen output of the program. It can be useful when you go back later and want to know what you did. In particular, check the "Model Report" near the start to ensure that the program is using the correct model.

    Now let's look at the current best topology. This is contained in a file called run1.best.current.tre. This file is updated every saveevery generations, so it is always easy to see the current best tree during a search. (Do not use this as a stopping criterion and kill the run when you like the tree though!)

  4. Open run1.best.current.tre in Figtree and examine the tree. (You may be able to double-click the file and associate .tre files with Figtree.)

exercise 3: inspect the final results of a run

Hopefully at least one of the search replicates has finished by now. Examine how the topology and branch lengths changed over the entire run. The run1.rep1.treelog00.tre file contains each successively better scoring topology encountered during the first search replicate. Note that this file can be interesting to look at, but in general will not be very useful to you. The default is for this file to not be created at all.

  1. Open run1.rep1.treelog00.tre in Figtree.
  2. Click through all of the trees.

    Note how the tree slowly changes over the run. We can also get other information from the treelog file.

  3. Open the run1.rep1.treelog00.tre file in a text editor.

    You will see a normal Nexus trees block. Each tree line includes comments in square brackets containing the lnL of that individual, the type of topological mutation that created it (mut = 1 is NNI, 2 and 4 are SPR, 8 and 16 are local SPR) and the model parameters of that individual. For example:

    tree gen1= [&U] [-10286.10914 mut=8][ r 1 4 1 1 4 e 0.25 0.25 0.25 0.25 a 0.922 p 0.394 ]

    If you scroll through and look at the mutation types, you will probably notice that a mix of all three topological mutation types were creating better trees early on, but the very local NNI mutations dominate at the end of the run. As a comment in each tree specification you will see the model parameters that were associated with each tree during the run. They are specified with a simple code:

    r = relative rate matrix 
    e = equilibrium frequencies 
    a = alpha shape parameter of gamma rate heterogeneity distribution 
    p = proportion of invariable sites 

The information that you really want from the program are the best trees found in each search replicate and the globally best across all replicates. After each individual replicate finishes, the best trees from all of the replicates completed thus far are written to the .best.all.tre file. When all replicates have finished, the best tree across all replicates is written to the .best.tre file.

The config files used here are set up to use a new feature of the program that collapses internal branches that have an MLE length of zero. This may result in final trees that have polytomies. This is generally the behavior that one would want. Note that the likelihoods of the trees will be identical whether or not the branches are collapsed.

When the two search replicates have completed, we can more closely examine the results.

First, take a look at the end of the .screen.log file. You will see a report of the scores of the final tree from each search replicate, an indication of whether they are the same topology, and a table comparing the parameter values estimated on each final tree.

There are two possibilities:

  1. The search replicates found the same best tree. You should see essentially identical lnL values and parameter estimates. The screen.log file should indicate that the trees are identical.
  2. The search replicates found two different trees. This is almost certainly because one or both were trapped in local topological optima. You will notice that the lnL values are somewhat different. The parameter estimates will be similar but not identical. The search settings may influence whether searches are entrapped in local optima, but generally the default settings are appropriate.

We can also evaluate and compare the results of our two search replicates with the PHYLIP package.

    We can use treedist to compare the two trees and see if they differ (and by how much).

  1. In the terminal window, type "cp run1.best.all.phy intree".
  2. In the terminal, run "treedist".
  3. Type "D" and press "Enter" to change the distance type to "Symmetric Difference". Then type "Y" to confirm.

    This will output an estimation of the "Symmetric Tree Distance" among trees to outfile. This distance is a measure of how similar the trees are, and is the number of branches that appear in only one of the trees. If the trees are identical, the distance will be zero. The maximal distance between two fully resolved trees is 2*(# sequences - 3).

    If the trees are different, we can calculate a consensus tree and see exactly where they agree. Note that in general you should choose the best scoring tree as your ML estimate instead of a consensus.

  4. In the terminal, run "consense". Confirm that you want to replace the previous output file and accept the default settings.

    This will generate a majority rule consensus. The majority rule consensus will show 50% support for branches that were only found in one tree, but it is not possible to show them all in a single tree.

    One final note: obviously you can visually inspect your final trees in Figtree or another tree viewer, but that is not a quantitative comparison.

exercise 4: constrain a search

If you looked carefully at any of the trees you've inferred (and know something about mammals), you may have noticed that the relationships are somewhat strange (and definitely wrong) in places. One relationship that this small data set apparently resolves correctly is the sister relationship of Whale and Hippo. This relationship (often termed the "Whippo" hypothesis) was once controversial, but is now fairly well accepted. If we are particularly interested in this relationship we might want to know how strongly the data support it. One way of doing this would be simply by looking at the bootstrap support for it, but we might be interested in a more rigorous hypothesis test such as a simulation based parametric bootstrap test or a Shimodaira-Hasegawa test. We won't go into those here, but in (Goldman et al., 2000) there is a detailed discussion of many of the available tests for comparing topological hypotheses.

The first step in applying one of the topological hypothesis tests is to find the best topology that does NOT contain the Whippo relationship. This is done by applying a constrained topology search. In this case we want a negative (also called converse) constraint that causes GARLI to search through tree space while avoiding any tree that places Whale and Hippo as sister.

GARLI allows constraints to be specified in a number of ways. We will do it by specifying the bipartition to be constrained using the "dot-star" format. This is a simple way of specifying a particular taxonomic grouping that uses a string of periods and asterisks, with one character for each taxon. All of the taxa represented with periods are on one side of the branch being specified, and all taxa represented with asterisks are on the other. For example, to constrain a grouping of taxa 1 and 2 in a data set of 8 taxa, the string would be "**......" (this is equivalent to "..******"). In this case the taxa we want to constrain are numbers 20 and 21, out of 29 total taxa.

  1. In the constrainedNucleotide directory, use a text editor to create a new (empty) text file named whippoNegative.con.
  2. On the first line, type in the string specifying the (Whale, Hippo) grouping (i.e., 19 periods, 2 asterisks, 8 periods).

    Note that GARLI will also take constraint trees. Constraints can either be positive (MUST be in the inferred tree) or converse (also called negative, CANNOT be in the inferred tree). The constraint type is specified to GARLI by putting a + or - at the start of the constraint string in the constraint file.

  3. Add a - to the beginning of the ....*** string that you just created, so that it looks like this: -...................**........
  4. Save the file.

    Now we need to tell GARLI to use the constraint. The garli.conf file in this directory has already been set up to be similar to the one we used during the unconstrained search earlier, so we only need to make minimal changes.

  5. In the constrainedNuclotide directory edit the garli.conf file and set constraintfile = whippoNegative.con.
  6. Set ofprefix = constrainedRun1.
  7. Save the config file.
  8. Start GARLI.

    Constrained searches can make searching through treespace more difficult (you can think of it as walls being erected in treespace that make getting from tree X to tree Y more difficult), so you may see that the two constrained search replicates result in very different lnL scores. When the run finishes, note the difference in lnL between the best tree that you found earlier and the best constrained tree. This difference is a measure of how strongly the data support the presence of the (Whale, Hippo) group relative to its absence. Unfortunately we can't simply do a likelihood ratio test here to test for the significance of the difference because we have no expectation for how this test statistic should be distributed under the null hypothesis. That is what the parametric bootstrap or a Shimodaira-Hasagawa test would tell us, but is well beyond the scope of this demo.

    Note the lnL difference between the best overall and best constrained trees.

exercise 5: perform an analysis using the GARLI web service

We have developed a GARLI web service that is backended by a powerful Grid computing system. You will use it for this exercise.

  1. Go to the create job page.
  2. At the top, enter an email address. This is how you will receive notifications about the status of your job.
  3. Enter a job name. This is completely arbitrary, but if you would like to help us track down problems, use your last name (e.g., Zwickl) as the job name.

    Since you now have more computing power available, you will increase the number of independent search replicates. Note that in the local run in exercise 1, you instructed GARLI to perform two independent search replicates (searchreps = 2). Now you will increase this number to 10. Each search replicate will run on a different processing node in our Grid system.

  4. Under General Settings, increase the Number of replicates to 10.

    We will be using the same data file that we used in exercise 1, but we need to upload it to the web server.

  5. Click on the Sequence data file field, and browse to the basicNucleotide directory.
  6. Select the murphy29.rag1rag2.nex data file.

    Now we need to set the model of sequence evolution for GARLI to use. We've already determined in exercise 1 that the best-fitting model under the AIC criterion is SYM+I+G, so we only need to change the equilibrium base frequency settings.

  7. Under Model Parameters > Non-partitioned Analysis > Nucleotide, change Base frequencies to equal.
  8. Fill in the CAPTCHA at the bottom of the page.
  9. Click the Create Job button.

    You should receive an email notification that confirms the job was submitted. The amount of time it takes to compute the job and get the results back will depend on a number of factors. It's okay to move on to the next exercise while you are waiting. When the job is complete, you should get another email notification containing a link to a page where you may download your results.

  10. On the job status page, enter your email address and the Drupal Job ID given in the email you received.
  11. Click the Get Job Status button.

    If your job is complete, a list of files should appear. The best_tree.tre file contains the tree with the highest likelihood score among the ten search replicates.

The web service is available for you to use any time. More features are available to you if you register for an account on the web site via the home page.

exercise 6: bootstrap analyses

It won't be practical to run bootstrap analyses during an in-class exercise, but you will find the output files from a GARLI bootstrap analysis of the Murphy29 data set in the bootstrapNucleotide directory. When performing a bootstrap analysis, GARLI will generate a number of bootstrap resampled data sets and perform a full tree search on each. The single best tree resulting from each replicate is written to a file named <ofprefix>.boot.tre.

GARLI does not currently calculate the bootstrap proportions or the bootstrap consensus tree from the set of trees found over the bootstrap replicates. An external program must be used for that, and good options are PAUP*, the CONSENSE program from the PHYLIP package, and a nice program called Sumtrees from the DendroPy package (it requires a Python installation, and is available at http://jeetworks.org/programs/sumtrees).

Here, we will post-process our bootstrap results using Sumtrees. This program can calculate a bootstrap consensus tree from the results of your bootstrap replicates. Also, an additional nice feature is that it can calculate the non-parametric bootstrap support for nodes and map these onto a target best tree generated from a non-bootstrapped ML analysis in GARLI. In this exercise we will (A) generate a bootstrap consensus tree using the nucboot.boot.tre file and (B) map the bootstrap support values onto a target best ML tree (using the run1.best.tre file that was created in exercise 3).

A. Make a consensus tree using the results of your bootstrap runs

  1. In the terminal, navigate to the bootstrapNucleotide directory, run the sumtrees program using the following command line:
    sumtrees.py --decimals=0 --percentages --output=myconsensus.tre nucboot.boot.tre
    This will calculate node support based on the bootstrap replicate trees, reported as percentages rounded to integers.

    This procedure will generate the majority-rule bootstrap consensus tree, including branches appearing in less than 50% of the trees. You will notice that some parts of the tree are very poorly supported, while others have high support. It is somewhat comforting that the parts of the tree that we know are resolved incorrectly receive low support. This is precisely why phylogenetic estimates MUST be evaluated in light of some measure of confidence, be it bootstrap values or posterior probabilities.

  2. Open myconsensus.tre in Figtree to see the consensus tree. Give a name to your support values in the pop up windows (e.g.: bootstrap), then check Node Labels and select your name from the Display drop-down menu.

B. Map the bootstrap support values onto a given tree (your best Maximum Likelihood tree)

  1. From the basicNucleotide directory, copy the file "run1.best.tre" into your bootstrapNucleotide directory
  2. In the terminal within the bootstrapNucleotide directory, run the sumtrees program using the following command line:
    sumtrees.py --decimals=0 --percentages --output=supportOnBest.tre --target=run1.best.tre nucboot.boot.tre
  3. Open supportOnBest.tre in Figtree to see the result

exercise 7: use checkpointing

One useful feature of the program is the ability to write "checkpoint" files during a run. If the run is stopped before finishing, either intentionally (e.g., need to reboot system, someone else needs to use system) or unintentionally (e.g., system crash, careless coworker closes the program), it may then be restarted from the checkpoint and resume where it left off. Note that a run restarted this way will give EXACTLY the same result as if it had never been terminated.

  1. In the checkpoint directory, edit the configuration file and set writecheckpoints = 1.
  2. Start GARLI.
  3. After the run progresses for a few seconds, stop it by pressing control-C, closing the window that it is running in or killing it some other way.
  4. Edit the configuration file again and set restart = 1.
  5. Start GARLI again.

    Note that it restarts from about where it left off. A run can be stopped and restarted with checkpointing as many times as desired. If you stop and restart it again, no further changes would need to be made to the config file because restart is already set to 1.

exercise 8: perform an analysis using a partitioned model

Partitioned models are those that divide alignment columns into discrete subsets a priori, and then apply independent substitution submodels to each. There are a nearly infinite number of ways that an alignment could be partitioned and have submodels assigned, so not surprisingly configuration of these analyses is more complex.

Note that although some models such as gamma rate heterogeneity allow variation in some aspects of the substitution process across sites, a model in which sites are assigned to categories a priori is more statistically powerful IF the categories represent "real" groupings that show similar evolutionary tendencies.

Running a partitioned analysis requires several steps:

  1. Decide how you want to divide the data up. By gene and/or by codon position are common choices.
  2. Decide on specific substitution submodels that will be applied to each subset of the data.
  3. Specify the divisions of the data (subsets) using a charpartition command in a NEXUS Sets block in the same file as the alignment.
  4. Configure the proper substitution submodels for each data subset.
  5. Run GARLI.

Note that detailed instructions and examples are available on this page of the GARLI wiki: Configuring a partitioned analysis

On to the actual exercise...

  1. In the basicPartition directory, open murphy29.rag1rag2.charpart.nex in a text editor. Scroll down to the bottom of the file, where a NEXUS Sets block with a bunch of comments appears. Notice how the charset commands are used to assign names to groups of alignment columns. Notice the charpartition command, which is what tells GARLI how to make the subsets that it will use in the analysis.
  2. Decide how you will divide up the data for your partitioned analysis. For this exercise it is up to you. There are a few sample charpartitions that appear in the datafile. If you want to use one of those, remove the bracket comments around it. If you are feeling bold, make up some other partitioning scheme and specify it with a charpartition. Save the file.
  3. Now we tell GARLI how to assign submodels to the subsets that you chose. Following is a table of the models chosen by the program Modeltest for each subset of the data. Look up the model for each of the subsets in the partitioning scheme that you chose. Don't worry if you don't know what they mean.
    sites rag1 rag2 rag1+rag2
    all GTR+I+G K80+I+G SYM+I+G
    1st pos GTR+G SYM+G GTR+I+G
    2nd pos K81uf+I+G TrN+G GTR+I+G
    3rd pos TVM+G K81uf+G TVM+G
    1st+2nd GTR+I+G TrN+I+G TVM+I+G
  4. In the basicPartitioned directory, open the garli.conf file. Everything besides the models should already be set up. Scroll down a bit until you see several sections headed like this: [model0], [model1]. This is where you will enter the model settings for each subset, in normal GARLI model format, in the same order as the subsets were specified in the charpartition. The headings [model0] etc MUST appear before each model, and MUST begin with model 0. For example, if you created 3 subsets, you'll need three models listed here. Open the garli_model_specs.txt file. This file will make it much easier to figure out the proper model configuration entries to put into the garli.conf file.
  5. In the garli_model_specs.txt file, find the models that appeared for your chosen subsets in the table above. For example, if I was looking to assign a model to rag2 2nd positions, the model from the table would be "TrN+G". Find the line that reads "#TrN+G" and copy the 6 lines below it. Now paste those into the garli.conf file, right below a bracketed [model#] line with the proper model number.
  6. Start partitioned GARLI.
  7. Peruse the output in the .screen.log file, particularly looking at the parameter estimates and likelihood scores. Note the "Subset rate multiplier" parameters, which assign different mean rates to the various subsets. Note that the likelihood scores of the partitioning scheme that you chose could be compared to the likelihoods of other schemes with the AIC criterion. Details on how to do that appear on the partitioning page of the garli wiki: Configuring a partitioned analysis

NOTE: The GARLI web interface now supports partitioned analyses as well. If you have time, try this new feature out on the web interface with the same dataset and partitioning scheme you used above.

exercise 9: perform an analysis using a codon model

Codon-based analyses are very, very slow. We won't do a complete search here, but will set up a configuration file for one and start it to get a feel for the speed. We'll tell GARLI to use a model more or less equivalent to the Goldman-Yang 1994 model.

  1. In the basicCodon directory, edit the configuration file. Some of it is already set up for this data set. We only need to set the details of the codon model.
  2. Set datatype = codon
  3. Set ratematrix = 2rate
  4. Set statefrequencies = f3x4
  5. Set ratehetmodel = none
  6. Set numratecats = 1
  7. Set invariantsites = none
  8. Save the file, start GARLI.

You'll quickly notice that the codon-based analysis is very slow, and you can kill it by typing ctrl-C.

exercise 10: further exercises

If you have your own data set of interest, now would be a good time to give it a try in GARLI. You can use the config file in the basicNucleotide directory as a template.

You might also try doing a constrained amino acid search that forces Whale and Hippo to be sister, since they are not sister in the ML amino acid tree. You can use the same constraint file as before, simply change the "-" at the start of the constraint string to a "+" to denote that it is a positive constraint. Comparing the likelihood of the constrained and unconstrained amino acid trees will give a measure of the support of the amino acid data against the Whippo hypothesis.