MEGA is a really convenient program for editing and getting a feel for sequence data. The latest version has incorporated tests among models of sequence evolution and also phylogenetic likelihood searches. Both of these are surprisingly rapid in comparison with other programs (partly because they are easy to implement in an analysis pipeline of alignment, annotation and trimming).
MEGA5 ranks models using the Bayesian Information Criterion (BIC) and automatically adds the parameter estimates for the best model as defaults for a likelihood search. Tamura et al. 2011 MBE 28: 2731-9 doi:10.1093/molbev/msr121 give some details of implementation, checks and comparisons with other programs. However, I am not seeing many papers where these features are used and I have heard some non-specific criticisms of the results (for example, that the optimal model chosen by jmodeltest and other programs differs from that chosen by MEGA5 - probably due to ranking by BIC rather than the Akaike criterion, which Tamura et al. 2011 find leads to overparameterized models). They have also cut some corners in likelihood searching that yields drastically reduced search times. For example, Close Neighbour Interchange (CNI) is the most severe tree changing algorithm, rather than the more drastic SPR or TBR, and they reject large steps in branch lengths between successive candidate trees.
If these improvements in efficiency are robust then this is very useful but are there some traps for the unwary in these conveniences?
Thanks Kaushik, MEGA still has quite a few quirks that puzzle me such as the inability to edit sequences in the data explorer.
On the Model Testing there is an unusual quirk in the calculation of the proportion of invariant sites (I).
For at least one dataset (Primates.nex example from the BEAST package) comparing models the HKY+G+I model was selected as equal best with TN93+G (BIC 11589.340, AICc 11393.088., LnL -5669.473). The HKY+G model (BIC 11603.516, AICc 11414.527, LnL -5681.198) ranked 5th on BIC, after TN93+G+I and GTR+G, and 6th on other criteria, after GTR+G+I. However, the estimated proportion of invariant sites for HKY+G+I was +I = 0.00 making this model indistinguishable from HKY+G, despite the reported difference in BIC, AICc and LnL. The same applies to the TN93+G+I and GTR+G+I models where +I = 0 was estimated but BIC, AICc and LnL values differed from TN93+G and GTR+G. A similar result can be seen in Fig. 1 of Tamura et al. 2011 MBE 28: 2731-9 MEGA 5 paper, where the JC and JC+I models have different BIC, AICc and LnL values despite an estimate of +I = 0.00 in the latter model.
How can an estimate of I = 0 produce a difference in the likelihood from an otherwise identical model without I parameter.
- Badges
- Science method
- Similar topics
- Biological Science
- Evolutionary Biology
- Phylogenetics
- Phylogenetic Analysis