How comparable are topologies derived from MrBayes and BEAST ?

[Steen]

I have been running the same data matrix with the same substitution models  and same partitions in both MrBayes 3.2 and in BEAST 1.7.5
I even use the resulting MrBayes topology (with polytomies resolved arbitrarily) as starting tree for my BEAST run.

However, I get a slightly different topology from MrBayes, compared to what I get from BEAST ...?
It looks to me like I am obtaining proper mixing in both my MrBayes run and my BEAST run, and ESS are also high.
I have rerun the same dataset multiple times.
When I run the dataset in MrBayes I continue to end up with topology A.
But when I run the same dataset in BEAST I get topology B.

The two or three clades that move around don't have high support in any of my two trees, but I would expect that the branching pattern in the two topologies ( A compared with B) at least would be matching?

Can anyone point me to what I am missing? Thanks in advance

Steen

[Jacek Kominek]

My guess would be that it's due to both programs having slightly different mixing strategies. Let's say MrB explores more of the branch-length space and BEAST explores more of the topology space, and the results differ because of that. I don't know if that would work, but if you really want to, you might try and set the mixing parameteres manually, and see what results come up.

If the volatile clades have low posteriors, then even if the sampling is good, you will tend to get incoherent answers simply because there is not enough phylogenetic signal to resolve them. In other words, if your data is conflicting, then you might take as many samples as you want from it - it still won't give you a coherent answer.

[Andrew Rambaut]

Dear Steen & Jacek,

The two programs will have different strategies for operators (moves) on parameters and it is possible one or other is less efficient at mixing but ultimately they should converge on the same result (if the models are precisely the same). If they never give the same result then it is not the mixing that is the issue. Also the high ESSs (I assume across all parameters?) suggest good mixing in both. The simple test would be to run them for longer to see if they converge. Another test would be to run them multiple times for both programs and check you get the same respective answer (use random starting trees).

Having ruled out inadequate mixing, it is likely that there is some difference in the model or prior. There are a lot of parameters and prior specifications so it might be difficult to completely match the two. What tree priors are you using? What molecular clock model are you using in MrBayes? etc.

If the clades that move don't have high support (you don't specify how low) then it is likely they are more influenced by the prior (i.e., the data is not being informative about those clades).

Best,

Andrew

--
You received this message because you are subscribed to the Google Groups "beast-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beast-users...@googlegroups.com.
To post to this group, send email to beast...@googlegroups.com.
Visit this group at http://groups.google.com/group/beast-users.
For more options, visit https://groups.google.com/groups/opt_out.

[Santiago Sanchez]

Hi Steen,

You can also run TreeAnnotator on MrBayes posterior trees, this way you can get a fully resolved  MCC tree to compare with your BEAST result.

Cheers,

Santiago

Santiago Sanchez-Ramirez

Ecology and Evolutionary Biology, University of Toronto

Natural History (Mycology), Royal Ontario Museum

100 Queen's Park

Toronto, ON

M5S 2C6

Canada

[Andrea Luchetti]

Hi to everybody,

I recently undergo the same problem: BEAST calculate a slightly different topology. 

I calculated a phylogeny with 61 congeneric termite species with a single outgroup species belonging to a closely related genus (adding further out group sequences lead to signal saturation).

In my resulting trees, I have a polytomy near the root of the tree (I obtained it using PhyML, RAxML and MrBayes) but BEAST groups part of the taxa of this polytomy in a monophyletic clade (pp=1.0) (of course all ESS are > 200; two runs that converged). 

The remaining of the topology is exactly the same as with the other methods. I also used Phylobayes, suppling the topology showing the basal polytomy (that obtained with RAxML), and it give a chronogram with some grouping very similar (bot not identical) to that of BEAST.

I also checked if the MCC from MrBayes tree list might be similar to the BEAST MCC, as suggested by Santiago, but it was still similar to the 50% maj_rule obtained with the MrBayes sumt (i.e., with the polytomy).

Node ages estimates are very close between BEAST and Phylobayes and, more important, the age of the ingroup is fully compatible the fossil data (the node of the ungroup was deliberately not calibrated).

Is it possible that some sort of long branch attraction may produce this results? Or there could be any other reason? I found few instances like that in literature, all relating about minor topology changes.

Cheers, Andrea

============

Andrea Luchetti

Dip. BiGeA - University of Bologna

MoZoo Lab

============

[Andrea Luchetti]