Mrbayes Amino Acid Model Probabilities And Acceptance Rates

Question

[Marcel]

Dear experts, I am trying to understand the meaning of the following output from a MrBayes run. Input alignment is aminoacid sequence and ProtTest said that Jones+G would be the most suitable aminoacid model for the dataset. In MrBayes I used the aminoacid prior 'mixedmodel' which takes all implemented models into account and decides on the fly which is best for the respective current tree. The acceptance rates for the cold chain of the two runssay this:

Acceptance rates for the moves in the "cold" chain of run 1:
With prob. Chain accepted changes to
31.58 % param. 2 (gamma shape) with multiplier
15.12 % param. 3 (topology and branch lengths) with extending TBR
30.95 % param. 3 (topology and branch lengths) with LOCAL
0.00 % param. 5 (amino acid model) randomly

Acceptance rates for the moves in the "cold" chain of run 2:
With prob. Chain accepted changes to
33.14 % param. 2 (gamma shape) with multiplier
15.01 % param. 3 (topology and branch lengths) with extending TBR 31.18 % param. 3 (topology and branch lengths) with LOCAL 0.03 % param. 5 (amino acid model) randomly

The way I read this is that the aminoacid model prior was not really suitable because the proposed models hardly ever got accepted.

Next, the Aaino acid model probabilities output looks like this:

Model - Post. Probability - Std.Dev.

---

Poisson - 0.001 - 0.000000
Jones - 0.012 - 0.017660
Dayhoff - 0.000 - 0.000000
Mtrev - 0.000 - 0.000000
Mtmam - 0.000 - 0.000000
Wag - 0.026 - 0.009183
Rtrev - 0.000 - 0.000000
Cprev - 0.956 - 0.026843
Vt - 0.000 - 0.000000
Blosum - 0.004 - 0.000000

The way I read this is that apparently the mixed model was pretty much useless because (i) in 95% of the cases Cprev was the best model, and (ii) strangely, Jones hardly ever got used although this was the model suggested by ProtTest. My interpretation is that I should try the next MrBayes run with Cprev as the selected aminoacid model.

Am I somewhat on the right path on how I interpreted this output or am I totally on the wrong track?

Any suggestions are greatly appreciated. Thanks a lot.

Marcel

Answer 1

[David W]

Hey Marcel,

I've not used the a/a models in MrBayes, so I can't offer you too much specific to that part of your question.

More generally, I think you are on the right track. The idea of the mixed-model is to be able to integrate all the other variables out across the posterior distribution for sequence evolution models, so you'd really want more changes to be accepted (the MrBayes manual suggests 10-70% for most proposals as a good ballpark). You can change the way in which propsal are made/accepted with propset but naively tweaking (the only way I know!) those settings is a very good way to get your chains to run for ever without converging.

If I were you, I'd run the analyses again using the model selected by prot-test and see if the topology/branch lengths are much different. That way have a predominantly cprev run and a completely Jones run and you'll be able to say your results are "robust with respect to the evolutionary model used" or some such.

Good luck ;)

[Marcel]

hey david, thanks a lot for the feedback and the reassurance. what you suggest makes sense to me ...