Need Advice On Performing Bayesian Model Selection Using The Mrbayes Tool.

Question

[Alice]

I'm just biologist and not an expert in Bayesian statistics, but try to learn how to properly and wisely use the bayesian approach. So, for my concatenated matrix (mtDNA, nuclearDNA + gaps in one of nuclear genes) I made partitioned analysis.

Fully resolved tree topology were yielded with matrix dissected on 9 partitions. To all of them were selected models of nucleotide substitutions (with jModelTest).

1. gene 1 1st nucleotide in codon
2. gene 1 2nd nucleotide in codon
3. gene 1 3st nucleotide in codon
4. gene 2 1st nucleotide in codon
5. gene 2 2nd nucleotide in codon
6. gene 2 3rd nucleotide in codon
7. gene 3 (rDNA)
8. gene 4 (rDNA)
9. gaps (as "standard" characters)

then

lset applyto=(1) nst=1 rates=gamma;
lset applyto=(2) nst=2 rates=equal;
lset applyto=(3) nst=6 rates=gamma;
lset applyto=(4) nst=1 rates=equal;
lset applyto=(5) nst=1 rates=propinv;
lset applyto=(6) nst=1 rates=propinv;
lset applyto=(7) nst=1 rates=gamma;
lset applyto=(8) nst=2 rates=equal;

After a while, I found Jeremy M. Brown's and Fredrik Ronquist's presentations, where they recommended not to choose models, but "let the [bayesian] analysis sample different models... (reversible jump)" and "If you use ModelTest or MrModelTest: Do not fix parameters in MrBayes"

Does it mean for me, that I did everything wrong? With such too detailed partitition I reсieve distorted tree topology, I guess. And can someone explain why is it need to find models by the Bayesian MCMC analysis itself?

Thank you for your attention. I will be glad to hear any comments regarding my analysis.

Answer 1

[Whetting]

Hi Alice,
First of all, I do not think you did everything (or anything for that matter) wrong in the analysis.
Even with the "lset nst= mixed" command you still need to specify different models of rate variation across sites. So I think it is ok to use ModelTest models in a Bayesian analysis. With the obvious caveat that a model is just that, and can over- or underestimate what is really happening.

As to why you would want MrBayes to test all the models was suggested in their paper (Huelsenbeck et al., 2004): "Current implementations of any of these criteria suffer from the limitation that only a small set of models are examined, or that the test does not allow easy comparison of non-nested models. In this article, we expand the pool of candidate substitution models to include all possible time-reversible models. This set includes seven models that have already been described."

I obviously do not know what tree you are running, but use common sense. The eyeball test (i.e. are things that should be together together ...) can go a long way.
Most importantly, make sure your trees have converged. I like to use AWTY (http://king2.scs.fsu.edu/CEBProjects/awty/awty_start.php) for that.

[Alice]

Thank you very much, I can breathe a sigh of relief. I haven't receive tree yet and look forward to SD of split frequencies reach the point of 0.01. Few days ago I met with AWTY, and obtained this picture (after 1 800 000 generations, 0.03 SD, ~150 sequences in analysis, relburnin=yes burninfrac=0.10 printfreq=1000 samplefreq=10000 nchains=4, 2 runs) http://www.flickr.com/photos/papilio_bianor/8496929487/

Because of my big sample I'm going to optimize analysis by playing with chain temp. and number of chains, hope this will help.

Оnce again, thank you for answer.