1
$\begingroup$

I have a couple of prokaryotic genes that I'm trying to build phylogenetic trees from.

I inputted them into IQTree, which, as part of the workflow, runs through multiple substitution models until it finds the one that seems to fit the data best.

My issue: quite a few of the genes ended up with eukaryotic_specific substitution models (such as Q.plant or Q.insect). Is this likely to bias my data? Should I re-run the analysis and remove the substitution models that are meant for different groups?

Thank you for your time!

Improve this question
$\endgroup$
5
  • $\begingroup$ That is a complicated question @Laura and is certainly worrying. I do know insect mutation matrices and compositional bias but cannot think why that would be confused for bacteria. What is the %GC of the bacteria of interest? Secondly, what is the data, I presume it is nucleotide data. $\endgroup$
    – M__ ♦
    Jun 9, 2022 at 2:19
  • $\begingroup$ Hello @M__ ! It's protein data, I forgot to mention. The GC is roughly 50-55%, with a few outliers (e.g. my outgroup has ~60% GC content). $\endgroup$
    – Laura
    Jun 9, 2022 at 8:23
  • 1
    $\begingroup$ Thanks only the protein bit matters ... and yes it is likely to make a different. Could you tell me whether it is using any other protein mutation matrix e.g. DayHoff, JTT or is Q.insect the matrix? It is fairly easy to generate a mutation matrix, rather than use an established one. $\endgroup$
    – M__ ♦
    Jun 9, 2022 at 9:06
  • 1
    $\begingroup$ Thank you @M__ . The program is running the data on a lot of amino-acid exchange rate matrices (including JTT, Dayhoff, and a lot of variants of each accounting for rate heterogeneity across sites iqtree.org/doc/Substitution-Models). It seems to pick the best model on something called Bayesian Information Criterion? (it also lists the best model according to Akaike Information Criterion and Corrected Akaike Information Criterion but it doesn't look like they influence later steps as far as I can see). Overall (with the variants) it goes through about ~700 models each time. $\endgroup$
    – Laura
    Jun 9, 2022 at 9:17
  • $\begingroup$ ... is there any chance the bacteria are insect symbionts @Laura? $\endgroup$
    – M__ ♦
    Jun 9, 2022 at 18:48

1 Answer 1

Reset to default
2
$\begingroup$

IQTree is a fashionable method for speeding up bootstrapping and used on ultra large datasets. What IQTree might be doing is rigidly fixing a model and parameterisation is not estimate via the ML. This might be how it achieves its performance.

In context, IQTree can use an establised mixed model, which is one approach to protein phylogeny. As a personal note, I don't favour it and come back to the reason later. It MIGHT also generate its own mixed model. There are not 700 established (published) amino acid matrices plus established mixed models, combined there are no more than 50.

There are around 30 protein mutation matrices, Dayhoff models (I honestly don't who Dayhoff is), one for insects which might be for mtDNA ... mtArt, one for mtDNA, mtREV I think and so on. Please check carefully the number of models it is evaluating.

Its easy to see how 4 or 5 combinations (or possibly permutations) of 30 become 700 models if it is generating its own mixed model.

Alternatively, what it does (mixed model or not) BIC and AIC do is given a tree estimate (prior) what's model is the best probabilistic fit. In your case Q.insect is cropping up, whether this is mixed model or not ... dunno.

What I don't know is how it calculates the mixed model, is it creating a new singular matrix by averaging (which is probably closer to the algorithm), or is it sequentially passing the data through first one model, then another etc ... Personally, I avoid mixed models for precisely this reason, I dunno what the actual matrix is, but accept it can provide strongest AIC/BIC.

It suggests either:

  • BIC/AIC results in an outcome that isn't biologically realistic .... OR
  • there has been some weird parallel/convergent evolution - this happens in parasites for certain

If you are using a single matrix not mixed model and Q.insect is cropping up - I would be concerned. If it is part of a mixed model, I wouldn't worry.

Is it checking 30 models (plus published mixed models) or 700 matrices??

If the final point was correct (biologically real result), you check it via blast - simply exclude bacteria and see if you get a plant of insect sequence as the top hit.

Ultimately, the uncertainty is the reason I avoid mixed models.

To answer the question, I'm not aware of a bacterial amino acid matrix. HOWEVER, you can select a mixed model. It might just be screening around 50 models via AIC/BIC and within that Q.insect is the most probable.

Summary The alternative is to look for support using a general amino acid matrix (for example JTT, LG or JTTDCMut).

Personally, I would be very cautious about proceeding with a matrix generated by insect genomes outside a mixed model for bacteria and in the absence of any biological support. FOR EXAMPLE, if these were insect symbionts - that makes sense.

Which ever approach is used a Bayesian approach is often required.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.