if they are othologous, the sub-functionalization model does not work since they are not paralogs... 

I would go for building the phylogeny as suggested by Gao, and then you can use PAML, DIVERGE, COMAP (From Dutheil and Galtier)  or other tools to test for selection. 

The problem of saturation Romain Studer is talking about might be a real issue, though the similarity between sequences (about 80% right?) shows that they are highly conserved for such a divergence time... So the saturation might not be an issue but few sites might be under selection...

did u test the similarity between the orthologous group, usually the ratio has a very wide range. so it is difficult to use the mutiple alignment on the group

@Gao Yes, I checked all the sequences have >80% similarity with each other.

it is very high, then it is good enough to do multiple aligment on it and then build the genetic tree

Yes, you can analyze selection. But keep in mind, that the presence of several gene copies usually reduces the negative selection pressure, especially if the two copies have split recently. This is because mutations that would be harmful if only one gene copy was present become less harmful if there is a second gene copy to compensate (see reference). It might also be helpful to first align the protein sequences and then recover the nucleotide alignment using the protein alignment as a guideline (if you are using the nucleotide sequences to estimate the time of divergence between sequences). 

Article Asymmetric and non-uniform evolution of recently duplicated ...

Yes, it would possible to test positive selection, with either orthologs and/or paralogs. There are different models using dN/dS in CodeML/PAML. I also built a tutorial for the detection of positive selection using the branch-site model: http://evosite3d.blogspot.co.uk/2011/09/identifying-positive-selection-in.html

My only concern is that your sequences are as old as the metazoan split time, so you might have saturation problem at the dS level (i.e. multiple changes at one synonymous site seen as only one change).

if they are othologous, the sub-functionalization model does not work since they are not paralogs... 

I would go for building the phylogeny as suggested by Gao, and then you can use PAML, DIVERGE, COMAP (From Dutheil and Galtier)  or other tools to test for selection. 

The problem of saturation Romain Studer is talking about might be a real issue, though the similarity between sequences (about 80% right?) shows that they are highly conserved for such a divergence time... So the saturation might not be an issue but few sites might be under selection...