Phylogenetic analysis, combining Bayesian and Maximum Likelihood

Question

[Adrian Pelin]

Hello,

I think I saw this answered before but cannot find the post.

I was wondering what is a publishable approach in using these 2 methods. What I used to do before, is run the analysis using both methods and employing the same model. Then I would pick the better supported tree and compare the tree side by side copying the support values.

I now have a problem since I have 100 taxa, and the trees are somewhat different. I like my bayesian tree, gives better support. Can I force this tree onto phyml and construct BS values based on the bayesian tree? Then I know my ML tree is the same as the bayesian one and I just have to copy the values over.

Adrian

Answer 1

[David W]

You tried to estimate something with two different and statistically valid methods, and got different results. Seems the right thing to do in that case is report _both_ results and discuss how the differences may have arisen and what the differing topologies would mean for th larger questons you are investigating.

You can ceratinly get your favourite ML program to calculate the liklihood of the optiam Bayesian tree (in phyml you'd use -u to provide a user-tree, and -o lr to optimise only the branch lengths and substituion mode. You could then to compare the likelihoods to see how strongly supported the differences between the trees are. (you can also recover the posterior probability of any clade you are interesting in from the Bayesian run by checking out the MCMC sample of trees with something like Dendropy).

[DG]

You can also generate bootstrap trees and do AIC tests with the two competing ML and Bayesian trees as well.

[Adrian Pelin]

So you would suggest generating 2 trees from ML and Bayesian, then feeding each tree as input for post prob estimation and BS? I end up with 2 trees, both with prob and BS values?

Is there an option to feed an input tree to mrbayes?

[David W]

You can't recover the clade probabilites for a given tree with MrBayes (so far as i know), but the answers are sitting in your posterior sample of trees (the .t files). Something like the sumtree utility from dendropy should do the trick

[DG]

What I have done in the past is generate bootstrap trees (lots) and do the AIC test on the topologies, so include the ML and Bayesian produced topologies in the same input file. You are really interested in the ML versus Bayesian topology but need lots of trees to properly conduct the AIC test from what I recall. It's been a few years since I have done it.

[Adrian Pelin]

Quick question. What does "-o lr" in that case do? I know it optimizes branch lengths and subsitution, but would that not change the tree? I mean the topology would be the same of course from the input tree, but BS support values would be computed on a tree where branch lengths and substitution rates were changed. Why not use "-o n"?

[David W]

So, I guess there are two slightly different approaches here. You can get BS support for the clades in the Bayesin tree without re-running anything (presuming the bootstrap trees themselves are saved). Just use denodropy or raxml (with options -f b -z [bootstrap trees] -t [MrBayes tree]) or possibly phyML but I don't know about that one.

Alternatively, you can see if how much "worse" the Bayesian consensus tree is, in terms of the ML value for that tree. If you do this, I would use the -o lr option, as in this way you getting the ML estimate of the model for the toplogy MtBayes found.

[DG]

I would suggest doing both personally, although it depends on just how different the overall topologies are, and in particular the clade(s) of interest. But both are fairly straightforward to do and are complimentary to one another and worth discussing in any publication where you have two competing topologies.