Thanks to everybody who kindly advised me on the problem.
Seemingly simulation is the only optio I may rely on. But still there is
some doubt which is more phylosophical.
As far as I understand, all algorithms of tree inference incorporate the
assumption that all sequences in the analysis are tips. Thus in case
sequences are nodes, ML, parsimony or NJ trees are likely to be false.
If I use a topology inferred with this assumption for parametric
bootstrapping/simulation, the results should be biased towards the
Is there a compromise?
Or I am totally wrong?
Anyway simulation is already practical and "publishable" :)