Partitioning analysis by both locus and codon position

[stefandahl]

Hi,

I'm a current Master's student and not very experienced with IQTree. I'm trying to run model testing on my data partitioned by locus as well as by first, second, and third codon position. I've tried partitioning the file as follows, as either DNA or codon (changing the prefix of the charpartitions to either DNA: or CODON:, example for the first locus):

#Nexus
begin sets;
charset ENSSHAP00000000220_1_mafft_BMGE_H07W10G05 = 1-462\1;
charset ENSSHAP00000000220_1_mafft_BMGE_H07W10G05 = 1-462\2;
charset ENSSHAP00000000220_1_mafft_BMGE_H07W10G05 = 1-462\3;
charpartition mine = DNA:ENSSHAP00000000220_1_mafft_BMGE_H07W10G05;
end;

The output file only seems to have one model for each locus, rather than selecting models for all three codon positions throughout the alignment. For example, for the first locus:

#nexus
begin sets;
charset ENSSHAP00000000220_1_mafft_BMGE_H07W10G05 = CODON, 1-462\1;
charpartition mymodels =
SCHN05+FU+I: ENSSHAP00000000220_1_mafft_BMGE_H07W10G05,
end;

Can someone help me figure out what I'm missing? I have attached the partition files (for codon and DNA) and the phylip alignment file in text format:

alignment.txt
partition_file.txt
partition_file_best_model.txt

Thanks,
Stefan Dahl

[roblanf]

Hi @stefandahl,

I think this is a fairly simple misunderstanding, but let me know if this helps.

Codon models (CODON) are a very different type of model to DNA models. To get a handle on the differences, I recommend this paper:

https://doi.org/10.1093/molbev/msn232

The reason you see one model per locus is because that's what codon models do - they model the codons themselves, not separate codon positions.

The long and short of it is that if you are interested in getting a good tree topology, and in finding a good way to model your sequences while doing so, you are probably fine to stick with DNA models (the vast majority of people do. Codon models are most useful when you are interested in selection, dN/dS analyses, etc.

For your case, you have already done all the hard work. To run a search for the best model what you do is start from the partitioning scheme with all codon positions modelled separately, and then use IQ-TREE to iteratively merge those codon positions together to keep improving the model.

How to do that is covered here:

https://iqtree.github.io/doc/Advanced-Tutorial

But the key part is just that you need a command line like this:

iqtree -s example.phy -p example.nex -m MFP+MERGE

One small technicality that's worth knowing is that you have three options where the -p goes:

• -p: linked branch lengths (partitions can evolve slower or faster than each other, but there's a single universal set of branch lengths)
• -q: equal branch lengths (all partitions evolve at the same rate; all branch lengths equal)
• -Q: unlinked branch lengths (every partition gets its own set of branch lengths)

Most studies find that -p is the best (e.g. https://academic.oup.com/mbe/article/37/4/1202/5673393), and most people don't bother with the other two. It's usually sufficient to cite studies like the linked one which show that linked branch lengths are fine. If you want to be super rigorous, you can run the analysis three times (one with each option), and choose the model with the best BIC score across all three.

I'll close this issue for now, but feel free to keep asking questions if you have them. If it's a new question, just open a new issue (that keeps the discussion forum clean and easy to navigate).

Rob