Hello Seunghyun Lee

Could you give a bit more detail on the analysis?

• How many taxa/genes/OTUs/terminal nodes are there?
• What model are you applying?
• Is it two MCMC chains you are comparing?
• What is the nature of the sequence data? (nuclear? amino acids? concatenated genes? mixed data types?)
• How complete is the matrix?
• When you do your bpcomp , what burn-in and subsample are you applying?
• What model did you apply in MrBayes, and are the high-confidence nodes in agreement with the Phylobayes nodes? When you say there are nodes below PP = 50, are you still rendering them as bifurcating nodes from the MB analysis? Why? (Given the lack of support, they should be presented as polytomies, unless you have a good reason to believe that node exists).

You could tell your bpcomp to rended nodes with PP below 50 as bifurcating, by adding the argument -c [threshold], for example -c 0.2 will render nodes down to a PP of 0.2, and present polytomies for nodes below that threshold. This should give you some insight into what confidence levels are like, but trees made like this are not valid to present as inferences, since you are presenting a result, with a statistics which say "the evidence does not support this": it is a contradiction. Equally, your MB tree with bifurcating nodes with PP below 50 is not valid.

On a theoretical point, what are your reasons for believing the MB inference but questioning the PB one? If you have a model which fits the data well, but it produces an unresolved tree, it doesn't mean you should favour another model which returns a resolved tree. Polytomies are not methodological "failures", but inferences manifesting from the data and the model. Given your maxdiff and ess stats, it sounds like your chains are agreeing on their uncertainty. Polytomies can be real results, and can be informative. This is especially the case when you have contrasting methods, as you have with PB and MB, which are highlighting that the inference is unstable. You need to work out if this is an aspect of the data, the model, or neither: depending on the context, polytomies can be true.

To investigate your contradictory results, you could

• look into where signal lies in the data by partitioning or subsampling the matrix in various ways - both across sites and across taxa - and reanalysing those matrices,
• compare models to identify which nodes are least stable, and try re-running analyses after removing taxa which seem to introduce the instability,
• look for biases from sequence saturation, long branch attraction, root placement etc,
• carry out model tests, or in PB cross-validate models (see the manual for how to do this),
• run the chains with the -s argument so that you can get more detail on the posterior statistics (see PB manual).

Hope that helps :)

I believe that the sequence data do not contain sufficient phylogenetic signal. So you have access to more sequences for your species.

Ah yeah, Patrice Showers Corneli has a more concise and important point! 5000 sites might not be enough data, and even if it was giving a resolved tree, that might be cause for concern because it may be an artefact of site specific bias. Also, if you are applying GTR, 5000 positions is probably not enough for the model to generate generalised exchangeabilities. If you are applying CAT, again, 5000 sites is probably not enough for the model to work with.

As Patrice says, these issues are dealt with by by adding loci / genes / independent sequences, so that site-specific biases are "diluted out" across the dataset, signal is amplified, and noise quietened.

Thanks Alastair. Well put. Sometimes 5000 sites can be enough to resolve a tree. It completely depends on the number of sites that are not either singletons (a site with just a change at just one taxa) or constant (no change at a site). I used a roughly 300 site alignment ( a short nuclear intron) to resolve a cone snail tree. This is a rare case where the variable sites told a consistent story across the alignment. The molecular tree conformed very nicely to morphological data. On the other hand an alignment built from much long sequences had such noisy data (conflicting sites) that it would not resolve as tidily as the short sequence. But these are rare I am guessing that you have an alignment with too few changes across taxa to resolve a tree.

Patrice Showers Corneli Alastair R Tanner

Thank you guys :)

My data contains 100+spp and 5000+bp widely used sanger-based DNA sequences. It would not generate additional data, that is my final dataset.

The question I put was "why some supporting values are SO different from those of MrBayes?"

I know MrBayes generally overestimate the nodal supports, but some basal nodes are too different (e.g., some nodes in MrBayes are pp>95/ while

That is a very good question. I'll look into it.

Oh that was fast. Here is my guess. Phylo-Bayes has a much more general set of models, called mixture models, than does MrBayes. Mixture models are more complex and therefore has less power than simpler models to resolve a branch. This is not a bad thing. To paraphrase John Tukey (I think): better to have the right answer with less support than a wrong answer that has high support. This is a common reality with large data sets. I suggest that that the Phylo-Bayes pp is probably more realistic.

What did the support value on that branch look like using Raxml, which is a maximum likelihood method rather than a Bayesian one. ML is generally less biased than Bayes for a number of reasons and Phylo-Bayes is almost certainly less biased than MrBayes. All depends on the priors you used for MrBayes.

Seunghyun Lee To the question: "why some supporting values are SO different?"

I think there are two major concept here.

Hi everyboby!

I am new user of Phylobayes.

I am trying to obtain a genome based phylogeny in order to know the phylogenetic position of an endosymbiont that I am working with.

Running Phylobayes 4.1c is not a problem, the command line is easy to run. The thing is that at least one of the independent runs of MCMC is stuck in a local maximum (maxdiff=1), even after 10 000, 20 000 and 30 000 cycles. Actually, checking the trace files (I ran 6 independent chains) it resulted that around the 500th cycle the loglik value of the chains were stabilized. Then, the topology of the consensus tree displays some polytomies even though the high support values for the nodes and the recovery of the main clades congruently with previous research references. I don't know how to deal with chains stuck in local maximums, I wonder if someone could help me with guidance to solve this issue.

Thanks in advance

hi Víctor Manuel Sosa Jiménez !

I also encountered the same problem, may I ask whether your problem was finally solved?