Dear Joanna,

I want to say that the branch lengths are drawn under a model which takes into account the number of substitution between your sequences and could also estimates those you cannot see (reversion)... The question is more now, is this model can estimate a reversion rate? For me, JC cannot do such things... so the branch length represents the distance between you sequence.

Best

Guillaume

Hello Joanna,

What made you choose the JK model?

The branch lengths represent maximum-likelihood estimates of the number of substitution events per site, including multiple substitutions and reversions, under the substitution model you use, and with the tree topology that produced the maximum likelihood among all the topologies that were explored.  This also takes into account uncertainty about the true ancestral nucleotide state at internal nodes.  The Jukes-Cantor model is not the best choice in most circumstances, because it assumes equal equilibrium nucleotide frequencies and that all possible substitutions are equally likely.  Recent versions of MEGA include an option to find the optimal substitution model to use for your data (similar to programs such as ModelTest), and I would recommend using this to choose the model.

Use GARLI for ML estimation. MEGA is not as good as GARLI in my humble opinion.

David Remington's answer on branch lengths is spot on. The number of differences observed among sequences in the raw data underestimates the actual number of substitutions that have occurred (even using Jukes-Cantor, which is the simplest possible model). The branch lengths reflect the number of substitutions implied by Jukes-Cantor distances.

Regarding the use of MEGA5+ for phylogenetic analyses...MEGA is an excellent choice for Maximum-Likelihood analysis. I haven't seen any evidence that the results are less reliable than any other package (although consistency among packages might depend on the dataset). In the release paper for MEGA5 (Tamura et al. 2011 Mol. Biol. Evol. 28:2731-9) the authors show highly similar results from MEGA, RaxML and PhyML (unfortunately not GARLI). In direct comparisons that I have done with PAUP* results are highly consistent, with no more difference than would be expected of any stochastic procedure (heuristic searching, bootstrapping). There are plenty of valid criticisms of MEGA, such as the limited search options, but reliability of results is not one of them. The implementation of ML in MEGA includes a number of computational efficiencies that make it much faster than other programs - this is hugely advantageous because it becomes feasible to do likelihood bootstrapping across large datasets.

The information Michael Cunningham provides about MEGA is really helpful, especially given MEGA's relative ease of use.