I am unsure if such a tool exists, but a person who would be good to contact about this is Allen Rodrigo at the ANU. He was involved in much of the early work about HIV phylogenetic trees.

I also have a method that can find the most conservative phylogeny for your data and removes sites that are likely to have undergone LBA. Let me know if you'd like to know more about it.

In BEAST or RAxML, you can specify monophyletic groups before you implement the phylogenetic analysis. Following the analysis, I believe you can compare the likelihood of trees using the program CONSEL. An example is Shen et al. (2015), "Enlarged Multilocus Data set Provides Surprisingly Younger Time of Origin for the Plethodontidae, the Largest Family of Salamanders."

Hello Brian,

What software are you using to infer your phylogeny?

When computers infer trees, they are always unrooted. Even defining a root post-hoc is just an aesthetic for human benefit, rather than any real change to the data which defines the tree (usually a newick string). Without a root being defined, software will still show a "rooted" phylogeny, often put the first taxon in the matrix as root, or sometimes it is just random. If this is the case, you need to define a root before assuming other problems with the tree. I would defer on assuming LBA until you are viewing your tree from a "human" rooted perspective.

If you have known ancestral nodes, you can reroot the tree using most softwares for viewing trees. I find the easiest FigTree  (link below). Paste the newick string into the window, select the branch you know to be basal/ancestral, and click reroot. From there you can then assess whether the tree is inside out or not.

If you then still appear to have LBA (or other biases), I'd do model tests on the matrix (I find IQTree the easiest to use and most informative, and I think it includes HIV specific substitution models for testing). If the matrix is large enough (as in, number of positions long), you might want to use the CAT model as applied in PhyloBayes, since this is the most effective current model for minimising long branch effects.

Best wishes,

Al.

http://tree.bio.ed.ac.uk/software/figtree/

http://www.iqtree.org/

http://megasun.bch.umontreal.ca/People/lartillot/www/downloadmpi.html

Thanks, Alastair.  It is usually not the whole tree that is misrooted or "inside out", but just one clade within the tree.  I am attaching here a tree that appears to have an example of the issue.  The HIV-1 M group subtype B epidemic began in Haiti and then there were hundreds of exports of the virus from Haiti to other parts of the world.  The sub-epidemic of subtype B introduction into South Korea happened in 1989 or 1990 and most likely did not come straight out of Haiti.  Subtype B spread into Trinidad and Tobago much earlier than 1990, but there was not much sampling and sequencing of viruses there until more recently, so there is a bit of a bias there for isolates in the post 2005 time period in comparison to USA/Europe where we have a lot of sequences from viruses sampled in the 1984 - 2000 time period.

In this tree, we might get the impression that the viruses from South Korea and Trinidad/Tobago originate near the "root" of the subtype B clade which I have labeled as Root 1 in red.  But it is far more likely, given all that we know of the epidemiology of HIV-1 subtype B, that the true root of the B clade is close the the green node I have labeled as "root 2".

When the tree is drawn as a radial tree like this, it gives us a bit of a different impression that when it is drawn as a cladogram.  And in either tree view, if we measure distances from tips to "root 1" we get different values than if we measure distances from tips to "root 2".  

Hi Brian,

Thanks for the clarification and further information.

In your rooted phylogeny, the Korea clade is, as you've noticed, on a long branch, and it is suspicious that the sister to this clade is also slightly longer than most clades. Also, the general shape of the tree is terminal long branches, with very short inter-node branch lengths. This is classic problem for inference - the substitutions which actually occur on inter-nodes will be swamped by substitutions happening along the terminal branches. How was this phylogeny inferred? What are your node support values like?

I would approach the possibility of LBA in the set in a few ways

• I would re-infer the tree using a variety of methods (and after model-testing the matrix), paying attention to difference in topology and node support.
• I would try sub-sampling the matrix so that each clade (of however you define them) is represented by a single taxa in the matrix, so that the matrix is only, say, 20 taxa, then reinferring. This diminished matrix will also be quicker to infer using Bayesian methods which can apply LBA mitigating models, such as CAT; it'll be easier to see if this approach will work on the full dataset, before using lots of computation resources.
• I would also analyse a dataset with the slightly-long-branched sister clade to Korea omitted, to see if the position of Korea changes with this group removed.
• You could also try re-coding the matrix, for example if it is nucleotide then converting to R-Y (purine-pyramidine), or if it is amino acids recoding to Dayhoff. Though this might be of limited use on your type of data - I am not very familiar with HIV sequence data...
• To return to your original question, yes there are ways of comparing the likelihood of competing fixed topologies, given a dataset (and a model), but I have never had to do them, so it might be worth waiting for another answer / doing some more reading about how to do that!

Hope that helps.

Brian, I haven't much experience in this particular issue (LBA), but I feel that the better way to deal with it is to include more taxa in your data. In particular, taxa that you may expect to be located just at the long branches would obviously help. However, adding taxa elsewhere in your tree may also improve character optimization at the nodes connected to the long branches.

As for comparing fixed topologies, I have fund the approximately unbiased and non-scaled bootstrap probability tests quite straightforward. The CONSEL suit (http://www.sigmath.es.osaka-u.ac.jp/shimo-lab/prog/consel/) provides tools to implement these tests.

Hope it helps

Thanks Leandro.  I have tried as few as 20 taxa up to more than 2,000 taxa and get the same result.  I have not looked at CONSEL, so I will check it out.

Thanks, Juan.  It is true that in most cases we do not know tthe true phylogeny and we are using methods to infer the most likely explanation of the data.  But HIV evolves in real time, and we very often have data sets where we know important details about the true phylogeny.  What I am seeking, is a way to rearrange the tree to fix what I know is an error, and then test the likelihood values of this tree vs the tree with the error, to ask if the two trees are really significantly different in likelihood or not. 

On thinking about this, it might be worth considering simply altering the Newick tree file and then using some kind of ML algorithm to correct the branch lengths once this has been done.

I just want to add: the CAT-like models for maximum likelihood are C10 to C60, which are supported in IQ-TREE. So you can use this when PhyloBayes is too slow. We also just published an approximation, which speeds up the analysis under such model without loss of accuracy: Article Modeling Site Heterogeneity with Posterior Mean Site Frequen...