evomics2019

Lab1: RAxML-NG Basics

Exercise 0 : Getting ready

1. Check input data:

cd /home/phylogenomics/workshop_materials/ng-tutorial/
ls

you should see roughly the following:

5.phy   dna.map    multi199.part  multi5.fa   prim2.part  prim.part  prim.phy.raxml.log  prot140.phy  prot21.fa.ckp  prot21.fa.out   prot21.part  README.md   terrace.part
bad.fa  fusob.phy  multi199.phy   multi5.map  prim.cons   prim.phy   prot140.part        prot21.fa    prot21.fa.log  prot21.fa.tree  rbcl.phy     terrace.fa

2. Run RAxML-NG binary without parameters

raxml-ng

this will show RAxML-NG version and short usage help:

RAxML-NG v. 0.8.0 BETA released on 11.01.2019 by The Exelixis Lab.
[...]
Usage: raxml-ng [OPTIONS]

Commands (mutually exclusive):
  --help                                     display help information
[...]

3. It is always a good idea to run alignment sanity check before starting the actual analysis:

raxml-ng --check --msa prim.phy --model GTR+G

This alignment is clean, so you should see:

Alignment can be successfully read by RAxML-NG.

4*. Run --check command again with bad.fa and examine error and warning messages.

Exercise 1 : Tree search

1. Infer ML tree with default parameters (10 random + 10 parsimony starting trees):

raxml-ng --msa prim.phy --model GTR+G --prefix S1

Check log-likelihoods for all 20 resulting trees:

grep "logLikelihood:" S1.raxml.log

Result:

[00:00:00] ML tree search #1, logLikelihood: -5708.961164
[00:00:01] ML tree search #2, logLikelihood: -5709.001321
[00:00:02] ML tree search #3, logLikelihood: -5708.928444
[00:00:03] ML tree search #4, logLikelihood: -5708.958315
[00:00:03] ML tree search #5, logLikelihood: -5708.932260
[00:00:04] ML tree search #6, logLikelihood: -5708.941449
[00:00:05] ML tree search #7, logLikelihood: -5708.959505
[00:00:05] ML tree search #8, logLikelihood: -5708.951658
[00:00:06] ML tree search #9, logLikelihood: -5709.022061
[00:00:07] ML tree search #10, logLikelihood: -5708.926872
[00:00:08] ML tree search #11, logLikelihood: -5709.016549
[00:00:08] ML tree search #12, logLikelihood: -5709.022648
[00:00:09] ML tree search #13, logLikelihood: -5709.009746
[00:00:10] ML tree search #14, logLikelihood: -5709.012081
[00:00:10] ML tree search #15, logLikelihood: -5709.017948
[00:00:11] ML tree search #16, logLikelihood: -5709.017067
[00:00:11] ML tree search #17, logLikelihood: -5709.030238
[00:00:12] ML tree search #18, logLikelihood: -5709.014300
[00:00:13] ML tree search #19, logLikelihood: -5709.018029
[00:00:13] ML tree search #20, logLikelihood: -5709.017251

2. Check topological distances between trees (so-called Robinson-Foulds or RF distance):

raxml-ng --rfdist --tree S1.raxml.mlTrees --prefix RF1

Result:

Average absolute RF distance in this tree set: 0.000000
Average relative RF distance in this tree set: 0.000000
Number of unique topologies in this tree set: 1

3*. Repeat steps 1 and 2 for fusob.phy.

Exercise 2 : Bootstrapping

1. Infer bootstrap replicate trees:

raxml-ng --bootstrap --msa prim.phy --model GTR+G --prefix B1

2. Map bootstrap support values to the best ML trees:

raxml-ng --support --tree S1.raxml.bestTree --bs-trees B1.raxml.bootstraps --prefix B2 

3. Open B2.raxml.support in the tree viewer of your choice.

Lazy "all-in-one" analysis

ML tree search, bootstrapping and branch support mapping can be performed with a single raxml-ng --all invocation:

raxml-ng --all --msa prim.phy --model GTR+G --prefix A1

However, this is not always possible/efficient for large datasets!

Exercise 3: Tree likelihood evaluation

The --evaluate command computes the likelihood of a fixed tree topology after re-optimizing all model parameters and branch lengths. We will use it to compare different evolutionary models:

raxml-ng --evaluate --msa prim.phy --tree S1.raxml.bestTree --model GTR+G --prefix E_GTRG

1. Repeat the above command for GTR+R4, GTR, JC and JC+G models. Don't forget to change the --prefix!

2. Check the log-likelihood scores:

grep "Final LogLikelihood:" E*.raxml.log

Which model yields the highest log-likelihood? Does it mean that this model should be preferred?

Parameter-rich models are more prone to overfitting the data. Thus, special criteria have been developed to compare models with different number of free parameters. In phylogenetics, AIC/AICc and BIC are most commonly used.

3. Check AIC/BIC scores for our runs (lower values are better):

grep "AIC score" E*.raxml.log

Which model should be preferred according to information theoretical criteria ?

Exercise 4: Protein data and ModelTest-NG

1. Check online help

modeltest-ng --help

Important options are:

• -i ALIGNMENT
• -d DATATYPE, where DATATYPE is nt (default) or aa

2. Run model selection for prot21.fa (protein alignment):

modeltest-ng -i prot21.fa -d aa

3. Run tree inference with the best-scoring model determined by ModelTest-NG.

4A. Re-run ModelTest-NG including the LG4M and LG4X models which are not tested by default.

Exercise 5: Partitioned models

Partitioned model is defined in a file, e.g. --model prim2.part.

cat prim2.part 

GTR+G+FO, NADH4=1-504/3,2-504/3
JC+I, tRNA=505-656
GTR+R4+FC, NADH5=657-898
HKY, NADH4p3=3-504/3

1. Re-run tree inference for prim.phy using partitioned model in prim2.part. Compare the results (log-likelihood and tree topology) to the Exercise 1.

raxml-ng --msa prim.phy --model prim2.part --prefix S2

Likelihoods:

grep "Final logLikelihood" S{1,5}.raxml.log

RF distances:

cat S1.raxml.bestTree S5.raxml.bestTree > S1S5.trees 
raxml-ng --rfdist --tree S1S5.trees --prefix RF5

Exercise 6 : RAxML-NG Web Server

Vital-IT/SIB kindly developed and maintain the RAxML-NG web server available here:

https://raxml-ng.vital-it.ch/

No registration or payment required. However, allocated computational resources are limited, so it is not a good option for large datasets.

If you have time, please feel free to play around with the web server (e.g., re-run Exercises 1 - 5).

Lab2: RAxML-NG Parallelization / Analyzing Large Datasets

Exercise 7: Alignment compression

Before analyzing a large dataset, it is highly recommended to convert alignment into RAxML binary format (RBA) using the --parse command:

raxml-ng --parse --msa fusob.phy --model GTR+G --prefix fusob

Doing this pre-processing step has two advantages:

• binary alignments are faster to load than PHYLIP/FASTA
• you will get estimated memory requirements and the recommended number of threads for this dataset

Exercise 8: Fine-grained parallelization

Use --threads option to re-run the same analysis with varying number of threads:

raxml-ng -search -msa fusob.raxml.rba -tree rand{10} -seed 1 -threads 1 -prefix T1
raxml-ng -search -msa fusob.raxml.rba -tree rand{10} -seed 1 -threads 2 -prefix T2
raxml-ng -search -msa fusob.raxml.rba -tree rand{10} -seed 1 -threads 4 -prefix T4

Compare the runtimes:

grep "Elapsed time:" T*.raxml.log

Which number of threads yields the lowest runtime? Which number of threads is "optimal"?

Exercise 9: Coarse-grained parallelization

Alternatively, we can parallelize across starting trees by running multiple RAxML-NG instances simultaneously:

for i in {1..2}; do (raxml-ng -msa fusob.raxml.rba -tree rand{5} -seed $i -threads 1 -prefix CT$i >CTlog$i &); done

This will start 2 RAxML-NG instances, each doing 5 independent tree searches on 1 CPU core. RAxML-NG will run in background, so use htop to monitor progress (look at per-core CPU load!).

Once all searches have finished, check the runtimes and compare them to fine-grained parallelization with 4 threads:

grep "Elapsed time:" CT*.raxml.log T*.raxml.log

Is it worth to use coarse-grained parallelization on this dataset?

Check the likelihood scores and pick the best tree:

grep "Final LogLikelihood" CT*.raxml.log | sort -k 3

Does it differ topologically from the best-scoring tree we obtained in Exercise 2? Do the likelihood scores differ? Why?

Exercise 10: ParGenes

Advanced topics

Exercise 11: Constrained tree search

Exercise 12: Multistate/morphological data