Phyml uses maximum likelihood optimization - perhaps the best method for reliable trees. Which option did you use in MEGA. Remember that different statistical phylogenetics programs produces tree estimates with different criteria so trees from different methods are often different.

Hi Patrice Showers Corneli . I used Maximum - Likelihood (Hasegawa-Kishino-Yano model) with Gamma-Distributed rates. This is identical to the one I used in PhyML (coded HKY+G). Oddly, the Maximum-likelihood tree and the Neighbor Joining trees generated by Mega are very similar (only the branch lengths differ).

Not sure why the Phyml score is different. Maybe a different optimizing method for each package. At any rate, I never expect the ML scores to be comparable across tree estimation software packages - to many numerical and algorithmic possibilities for the difference. What really matters is the ML score you get in each package. The comparison should used to compare models of sequence evolution to make sure the model is right (I assume you fit various models and found HKY+G had the highest likelihood). Comparisons of alternative tree topologies is another reason to compare likelihood scores so that the best tree is the one with the highest likelihood (although you should consider alternatives with small differences [ ≤ 2lnL units] to fit your data equally well).

Regarding the branch lengths, those for the ML tree should be the most accurate compared to any made by other methods. It is actually the branch lengths that are optimized in an ML analysis with the topology that best fits those branch lengths chosen as the one that best fits the branch lengths. Of course this is only true is you have chosen the best possible model of sequence evolution for your analysis,

The branch lengths are extremely important to get right because of long branch attraction and for the accuracy of any clocks that you use to find a divergence date.

Different methods will always give different trees, if the alignment is good and by that i mean there is phylogentic signal in the alignment then tree topologies will be similar regardless of the method or software package. We are inferring creating hypotheses as such slight differences should not impact on your interpretation. I would recommend you try ML, Parsimony and Distance methods to see if;how topologies change. Likewise Bootstrapping will help you decide if topologies have changed significantly. Keep up the good work :-)

Hi all,

Thank you very much for your comments. The parameter options are indeed different between MEGA and PhyML (even RaxML), so doing "equivalent" runs between software packages might be difficult.

Upon doing a rerun of PhyML (with more sequences) and following the recommended parameters of jModelTest2 (no approximation of likelihood test ratio & no bootstrapping), I finally got a tree similar to the one generated by MEGA7. They're similar but not identical, though that's acceptable since MEGA does not have the recommended substitution model (F84). I'm trying to check what changed. Running with or without bootstrapping (in PhyML) doesn't seem to have an effect on the the phylogeny, so I'm looking into alignment trimming and the other additional sequences I used.

Good luck to your. Your question was most intelligent. Feel free to ask for any further advice or opinions.

If the aligment is accurate the all of methods give you almost similar results except minor difference which is always there