Disallow LG4M+F and LG4X+F

[thomaskf]

Hi Hector, I now do recall that Olivier talked with me about this a while ago. I agree that LG4M+F and LG4X+F don’t make sense. I think we should better make an error message, instead of just warning. The error message needs to be informative.

@thomas Wong Can you make a pull request about this?

--
Minh Bui, Associate Professor
Leader, Computational Phylogenomics Lab
School of Computing, Australian National University
Hanna Neumann Building, Room 4.25
145 Science Rd, ACT 2600, Canberra
http://anu-phylogenomics.github.io

From: Hector Banos Cervantes <Hector.BanosCervantes@csusb.edu>
Date: Thursday, 16 October 2025 at 10:26
To: Thomas Wong <thomas.wong@anu.edu.au>, Minh Bui <m.bui@anu.edu.au>
Subject: Comment about model LG4X

Hi Minh and Thomas,

I hope all is well! I wanted to ask for a favor. Olivier Gascuel contacted me, saying that he thinks the use of the model LG4X when using the +F is not an accurate reflection of the model (that is, when using the model LG4X+F). Therefore, I wanted to see if a flag can be run when such a model is run, just so the user understands what the model is doing.

I guess is that it could be confusing for the user when the flag +F replaces the stationary frequencies and when it is added. For example, LG+F has a single frequency class, but C60+F has 61, so it is confusing what LG4X+F would do. I completely understand that from reading the IQTree output, one could deduce that, but maybe just printing a warning saying that in LG4X+F has a single frequency class (which does not have the design features of the original LG4X model). If you think this is not relevant or needed, I would understand, but I just wanted to pass the message along. Olivier was suggesting just removing the option of LG4X+F, but I do think that since people have used it, it should stay there for legacy purposes. Nonetheless, I leave it up to you to whatever you think must be done.

Kindest regards,
Best,
Hector Banos
Assistant Professor
CSUSB - Department of Mathematics

[roblanf]

Hi All,

Actually I think a warning would be preferable here, though I'm definitely open to discussion. Here's my thinking. People make 'inappropriate' combinations of models like this all the time. For example, C10-C60 were estimated with a POISSON matrix, but they are much more commonly combined with the exchangebailities from LG or WAG. Equally, LG was estimated with its own frequencies, but it's very common to re-estimate just the frequencies, conditioned on the exchangeabilities (though this seems less egregious, because at least the number of frequency classes matches up). My point is that sometimes these combinations are OK, sometimes they may not be, and often this is a legitimate research question.

I think we end up with three types of model in this sense:

1. A model where all parameters were co-estimated, so you end up with a coherent Q matrix (or mixture of them). E.g. LG+FU, LG4X, POISSON+C10, etc.

2. A model where some parameters are being re-estimated or combined, but the number of classes of what is being re-estimated remains true to the original in both directions. E.g. LG+F, LG+FQ, GTR20+F.LG (frequencies from LG matrix), LG4X+FMIX(4 classes), GTR20+C10

3. A model where some parameters are being re-estimated or combined, but at least in one direction the number of classes of what is being re-estimated differs from the original. E.g. LG+C10 (LG was estimated with one frequency class), LG4X+F (LG4X was estimated with 4 frequency classes).

My preference would be to catch each of these cases and print a quite verbose warning message with each. The reason is that by doing this we can allow the research community to do simulations and figure out when and why things can be bad or not so bad. For example, Hector's excellent paper showing that adding +F to C10-C60 models can be bad is a very important contribution. If we disable that possibility, it will be much harder for people to figure this kind of stuff out.

Thoughts?

Rob