How can I obtain per site log-likelihoods for a topology against an alignment without optimising branch lengths? (which software can I use to do this)

Per site log-likelihoods can be used for various tests of whether a topology fits the data in a multiple sequence alignment well using e.g. the Shimodaira-Hasagawa test, the AU test etc.

RAxML, Garli and FastTree and other software packages can calculate per site log-likelihoods given a tree, alignment and substitution model but all will automatically optimise branch lengths first.

I would like to test a tree in which a pair of sequences (A and B) are monophyletic, where in the maximum likelihood topology they are not. My problem is that when I infer a tree with a constraint requiring A and B to be monophyletic, the edge length is set to the minimum within the allowed precision. But having an edge length so close to zero does not make biological sense if the tree is describing a monophyletic clade for A and B: the minimum length of the edge must be equivalent to one substitution or character change, else A and B are in a polytomy containing clade with other sequences which is not the biological scenario I'd like to test (in that case there would be ambiguity regarding the relationship of A and B to other groups emerging from that polytomy).

So . . . I'd like to edit my constrained tree and increase the edge length to the A and B clade to the minimum one substitution length, then perform the AU test against the per site log likelihoods for that topology. But the above software optimises the edge lengths, resetting the edge length to A and B to effectively 0!



1 Answer 1


If you have the budget for it, I highly recommend Geneious for topology work. It was a real saver for me on an epi paper last year.

Anyway, they allow for testing per-log likelihood without adjusting edge length. If you want to throw up some dummy data somewhere I can throw up an example. I have no connection or stake with the company, just happy with it.

  • $\begingroup$ It turns out the tests I mentioned (SH, AU) are strictly tests of topology, not branch length. So choosing arbitrary branch lengths as I proposed would bias the test (according to Prof Alexandros Stamatakis; I'm assuming choosing an arbitrary topology does not cause a similar bias, because that is what the test is designed to assess). Thanks for answering the question though. $\endgroup$
    – DaveUU
    Nov 21, 2014 at 16:14
  • $\begingroup$ Yep, that's why I mentioned without "adjusting" or scaling edge length. I don't think you would ever want assign an arbitrary branch length, or at least I can't think of a reason to. $\endgroup$
    – Atl LED
    Nov 21, 2014 at 16:45
  • $\begingroup$ Because a bunch of adjacent internal edges of length zero makes a polytomy but a bunch of adjacent internal edges of non-zero length does not make a polytomy - it makes a series clades with events between them that may be inferred from the sequence data in the same way a particular topology can be. These are different biological hypotheses that you might want to test, in the same way assigning arbitrary topologies are different biological hypotheses. $\endgroup$
    – DaveUU
    Nov 21, 2014 at 17:07

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .