Maximum Likelihood (RAxML)

Previous class check-up

• We studied the algorithms for maximum likelihood and focused on the main contributors to good or poor performance

Learning objectives

At the end of today’s session, you will be able to

• explain in details the RAxML inference methods
• use RAxML software

Pre-class work

Read the papers :

• RAxML1
• RAxML2

In-class discussion

Individual work (15 minutes)

Instructions: Based on the RAxML papers you read, write down (somewhere) the following questions:

1. Starting tree
   • How does RAxML generate or accept a starting tree?
   • Why does starting tree choice matter for ML search?
2. Model of evolution
   • Which models are implemented in RAxML?
   • How are models chosen in practice?
   • What are assumptions or limitations of these models?
3. Data quality
   • What does RAxML assume about the input alignment?
   • Are there warnings or guidance about poorly aligned or incomplete data?
4. Convergence
   • How does RAxML search tree space?
   • How does it decide when it has found the ML tree?
   • What steps ensure that results are reliable (e.g., multiple searches, bootstrap)?
5. Optional extra credit:
   • Identify one tip or recommendation from the paper for improving runtime or accuracy.

Whole class work (15 minutes)

Let’s discuss the questions together!

Take-home message: always read carefully the paper and the documentation of any phylogenetic method you use

Create your own cheatsheet with description, strengths, weaknesses, assumptions and user choices (and other things you find relevant).

Software

The repository for RAxML has installation instructions here.

I click on Download OSX/macOS binary (x86 and ARM) and I download raxml-ng_v1.2.2_macos.zip. I put this folder in my computer (I have a folder software), and inside it is the executable raxml-ng.

I will copy this executable into /usr/local/bin so that I can call it from wherever in my machine, but you can also use the whole path.

cd software/raxml-ng_v1.2.2_macos/ ## you need to modify to your own path
ls ## check the raxml-ng executable is there
cp raxml-ng /usr/local/bin

We will use the primatesAA data again. For now, we will use the alignment from MUSCLE, but it would be interesting to compare to other alignments.

First, we will check the alignment:

raxml-ng --check --msa primatesAA-aligned-muscle.fasta --model LG+G8+F

And then we find the ML tree:

raxml-ng --msa primatesAA-aligned-muscle.fasta --model LG+G8+F

Version 2.0 or higher has now automatic model selection (see here).

RAxML produces the following output files:

Best ML tree saved to: primatesAA-aligned-muscle.fasta.raxml.bestTree
All ML trees saved to: primatesAA-aligned-muscle.fasta.raxml.mlTrees
Optimized model saved to: primatesAA-aligned-muscle.fasta.raxml.bestModel
Execution log saved to: primatesAA-aligned-muscle.fasta.raxml.log

We can do a quick and dirty plot in R:

library(ape)
tre = read.tree(file="primatesAA-aligned-muscle.fasta.raxml.bestTree")
plot(tre)

It is not enough to estimate the ML tree, we also want to do non-parametric bootstrapping. We call this non-parametric bootstrapping because we resample our alignment columns with replacement and rebuild trees many times to see which clades are consistently recovered.

We use the flag --all to run both ML and bootstrap:

raxml-ng --all --msa primatesAA-aligned-muscle.fasta --model LG+G8+F --bs-trees 10 --prefix primatesAA-aligned-muscle-raxml-boostrap

We are only doing 10 bootstrap replicates for the sake of time, but we should always try to do at least 100. We need to choose a new prefix because it doesn’t let us overwrite the previous raxml output files.

The output file we are interested in is:

Best ML tree with Felsenstein bootstrap (FBP) support values saved to: primatesAA-aligned-muscle-raxml-boostrap.raxml.support

We can do a quick and dirty plot in R:

library(ape)
tre = read.tree(file="primatesAA-aligned-muscle-raxml-boostrap.raxml.support")
plot(tre)
nodelabels(tre$node.label)

First, we note that the tree does not seem to be rooted correctly.

very important

Maximum likelihood methods are uncapable of inferring the place of the root. We always have to root the estimated tree afterwards with an outgroup.

library(ape)
tre = read.tree(file="primatesAA-aligned-muscle-raxml-boostrap.raxml.support")
plot(tre)
nodelabels()

rtre = root(tre, node=33, resolve.root=TRUE)
plot(rtre)
nodelabels(rtre$node.label)

How do we interpret those numbers?

Common mistake in final project: forget to root the estimated trees!