The other key feature to pay attention to, when selecting a model, is that for nearly all DNA and protein data, the addition of rate variation between sites (gamma distribution plus invariant sites is usually the best option) makes a larger contribution to selecting a good model than the model itself (F84, Kimura, GTR, etc).   This is because almost all sequences we are interested in are under selection pressures, not so much because different sites experience more mutations.  Selection pressures are not well modeled but including a difference in rate of evolution in different sites greatly helps to compensate for this.

Well, theroeticians might argue forever about the virtues of models... But the sad true is that ALL models of nucleotide substitution are simplistics toys. Useful if you know what you are doing, granted.

The idea is to pìck up a model that BEST fits the data. That is, that allows the algorithm you are using to choose a tree with the highest likelihood. This means you should select a reasonable model BEFORE running the tree-building program.

Fortunately there are several practical ways to do this. I like this one for its simplicity and speed: http://www.hiv.lanl.gov/content/sequence/findmodel/findmodel.html

With most sequence data you will find that GTR performs optimally, or at least is among the top scores. However, you should run Findmodel (or a similar program) for every new set of sequences.

Hope this helps!

The other key feature to pay attention to, when selecting a model, is that for nearly all DNA and protein data, the addition of rate variation between sites (gamma distribution plus invariant sites is usually the best option) makes a larger contribution to selecting a good model than the model itself (F84, Kimura, GTR, etc).   This is because almost all sequences we are interested in are under selection pressures, not so much because different sites experience more mutations.  Selection pressures are not well modeled but including a difference in rate of evolution in different sites greatly helps to compensate for this.

Check this good article:

Molecular phylogenetics: state-of-the-art methods for looking into the past

Simon WhelanPietro LiòNick Goldman(2001) Trends in Genetics Vol.17(5):262–272, doi:10.1016/S0168-9525(01)02272-7

The article cited by Francisco Henao-Diaz above is nice, and the data used for the sample analysis is from:

Brown, W.M. et al. (1982) Mitochondrial DNA

sequences of primates: tempo and mode of

evolution. J. Mol. Evol. 18, 225–239

I have attached the alignment (a bit less than 1kb of mitochondrial genome data) here in FASTA format, in case you want to use it to "test drive" your methods and compare to theirs. 

Hi Ahasan,

   I'd suggest you read review paper by Sullivan & Joyce 2005  and book by Ziheng Yang 2006 for some general ideas about DNA evolution models in phylogenetics.

Jack Sullivan and Paul Joyce. Model Selection in Phylogenetics. Annual Review of Ecology, Evolution, and Systematics. Vol. 36: 445-466 (2005) DOI: 10.1146/annurev.ecolsys.36.102003.152633

http://www.annualreviews.org/doi/abs/10.1146/annurev.ecolsys.36.102003.152633

Ziheng Yang. Computational Molecular Evolution, (2006) Oxford University Press

Hope this help.

Hello Ahasan.

Like other people have already correctly pointed out, it is prudent to run your data through a "model finder" algorithm in order to find with substitution model best fits your data.

Most of the times, you will land on a GTR model. A GTR model is one that is time reversible (simply put, it assumes that a base can change and than change again to the original one. e.g. A -> T and then T -> A). GTR also assumes separate rate parameters for each type os nucleotide substitution and assumes different frequency parameters for each nucleotide type. You could say that it is "the most independent substitution model", since the parameters are "free" between themselves.

You can further elaborate this model assuming that some portions of your data are invariable and/or have distinct variability rates.

I'd also suggest the same paper suggested by Chengmin

Jack Sullivan and Paul Joyce. Model Selection in Phylogenetics. Annual Review of Ecology, Evolution, and Systematics. Vol. 36: 445-466 (2005) DOI: 10.1146/annurev.ecolsys.36.102003.152633

http://www.annualreviews.org/doi/abs/10.1146/annurev.ecolsys.36.102003.152633

Hope this helps.

Good luck,

JP

All of the generally used models of sequence evolution are simplifications of the GTR model, thus it is not surprising that the GTR model is most often chosen in the "model selection" programs others have mentioned... For example HKY85 constrains the number of parameters by making that assumption that all types of transitions occur at the same rate, and that all possible transversions occur at the same rate... Jukes-cantor (1969) is a further simplification that assumes all possible mutations occur at the same rate. Thus all of these other models only differ from GTR in the degree of parameterization... And to prevent over-parameterization, many methods of comparing model fit (see JModelTest) will often penalize for the number of parameters- because a model with more parameters will generally offer a better fit.

Other aspects considered in your sequence evolution model is the degree of rate heterogeneity in your dataset, which we often assume to be gamma-distributed. The consideration of this distribution is something you will also consider when selecting the model that you are assuming in your phylogenetic analysis... And you will see in the literature as "+G" (e.g. GTR+G). To speed computation, this distribution is often "binned" (rather than treated as continuous) and so one parameter you will often see as manipulatable in phylogenetic programs is the number of bins. You will also often see that models may assume some proportion of your alignment to be invariant (represented as "+I"; e.g. GTR+G+I).

Reconstruct the evolutionary history of sequences, is not equal for all regions within the genome, each has its own pattern of evolution , so in recostruccion phylogenies these features must be considered , if the recostruccion is based in ML or genetic distances , where models are used for recostruccion . One option is to select the model using test models. David Posadas develop ModelTest and jModelTest software that allows selection of models with different modalities such as : distribution range and invariable sites .to fiind the best modei it use  diferet selection criterias such as hLRT and dLRT, AIC and BIC  This type of analysis can also be run in MEGA 7 in selection of best model of nucleotide and protein

https://www.researchgate.net/profile/Keith_Crandall/publication/13370131_MODELTEST_testing_the_model_of_DNA_substitution/links/0fcfd50ace03771cc6000000.pdf

Article Modeltest: testing the model of DNA substitution

thanks all!