My preference is to use protein sequences for multi-alignment and then replace each amino acid with the original codon. We build trees with the codon alignment. For deep branches, you more rely on the 1st and 2nd phases and for shallow branches, you more rely on the 3rd phase. At least to treefam, as I have tested, this strategy gives the best trees. There is also a review about what alignment to use when constructing trees (I cannot find it now). The conclusion is similar: codon alignment yields better trees for both deep and shallow branches.

Ah, you can have a look at my PhD thesis about an evaluation. Table 6.2.

Hi, if your sequences are not too divergent, you can use directly codon models in order to reconstruct your tree. PAML http://abacus.gene.ucl.ac.uk/software/paml.html can be usefull for doing this, it is perhaps not the most efficient program for phylogeny, but allows you to test for different codon models, have a look to the manual in order to know the corresponding number of degrees of freedom etc...

an other tool is https://www.nescent.org/wg_garli/Main_Page but I have almost no experience with it.

I think that if you can defend that there is no saturation in the rate of synonymous mutations, using codon model is the best option.

A way to check it would be to compare the tree support you obtain with codon model, or aminoacid model.

good luck!

Yes, use DNA if the protein sequences are very conserved. One advantage of using DNA sequences not mentioned so far is that you can infer a substitution matrix from the alignment, even with a smallish amount of data (like a single gene). You can't really do that for protein sequences unless you have a concatenated alignment (lots of data), so with proteins you often have to use pre-baked substitution matrices like WAG, LG, etc.

I think for older evolutionary splits, protein sequences will definitely be more informative, though - DNA will be too noisy. I suppose DNA could have more information if you were using a codon model but people don't seem to do that. Column-by-column, the signal in the four possible DNA bases will probably get saturated more quickly.

I found http://www.bio.net/bionet/mm/methods/1998-January/064055.html a good overview of why using which blast program. To sum it up, you would use Blastn with very similar sequences (think of two different strains of the same bacteria, or two individuals of the same organism) or to find non-coding regions.

Blastp is better for more distantly related proteins (since protein sequence is more conserved than DNA sequence) and it is faster than blastn because of a smaller database size.