ModelFinder in IQ-TREE3 could not find the best-fit model

[jeremiasivan]

Dear IQ-TREE Development Team,

I have the following FASTA alignment with four species: window_1_025_2.fa.zip

I noticed that IQ-TREE2 and IQ-TREE3 chose different models for the respective alignment: IQ-TREE2 chose TIM2+F+G4 (iqtree2.log), while IQ-TREE3 chose TPM2u+F+I (iqtree3.log).

So, I did the following:

1. Fix the seed number to be 12345
2. Fix the tree topology to be iqtree2.treefile (iqtree2.treefile.zip)
3. Fix the branch lengths with -blfix
4. Fix the model to be TIM2+F+G4

In this case, IQ-TREE2 and IQ-TREE3 returned the same results, with tree log-likelihood of -30169.352:

• IQ-TREE2: iqtree2_wtree_blfix.log
• IQ-TREE3: iqtree3_wtree_blfix.log

However, when I did not fix the model (but fixing other parameters above), IQ-TREE3 selected TPM2u+F+I again even though it resulted in lower tree log-likelihood (-30182.924) than TIM2+F+G4: iqtree3_wtree_blfix_nomodel.log

I noticed that in iqtree3.log, the likelihood calculation for TIM2+F+G4 model was skipped. Is there any change in how ModelFinder selects the best nucleotide model?
Thanks!

Regards,
Jeremias

[roblanf]

Hey @jeremiasivan!

Hard to see what's going on exactly, because the logs for IQTREE3 don't have the ModelFinder details (not sure why). They should be in the .iqtree files though.

Can you send the .iqtree files?

Rob

[jeremiasivan]

Oops, I forgot to double check the log files before uploading.

Here are the .iqtree files: iqtree3_iqtree.zip!

Jeremias

[evolbeginner]

Hi there,

Thx for the Q! While I am not an expert on this issue, based on my limited knowledge, in iqtree3 that the model TIM2+F+G4 is skipped could potentially be b/c TIM+F doesn't seem to have a very good AIC/BIC during a first evaluation, although with +G the subs model displays a much better fit to the data. If it really is so, then the idea behind it would be to speed up model selection. If needed, perhaps you can try sth like -mset K80,GTR,TIM2 -mrate +E,+G,+I,+G+I to force it to run a "complete" model test? Perhaps other experts may also want to help.

best,
sishuo