Thanks for your reply Artur, I suppose that was the question I had in mind, whether or not I could (or should) try to assemble a tree from a mixed data set (nt and aa) for different taxa. Since you can't do that I've just saved myself the time and agony trying to achieve the impossible.

I have to keep reminding myself that I am not trying to resolve the phylogenetic positions of all the vertebrata with this one gene, although it to a large degree does, right down to separating cartilaginous fishes from ray-finned fishes, monotrems from marsupials and Xenarthra from Afrotheria, and so on. I'll just have to keep the MCMC analysis going for longer to see if that doesn't shake down the more basal species in the nucleotide analysis. 

Another question: is a molecular clock gene defined as one that is subject to neutral or near neutral mutation rates, or one that is under purifying selection. A duplication independent homologous gene must by definition be under negative selection or else it would not maintain its homology across biological time and space. I've read authorities such as Zuckerkandle, Nei, Kumar and sundrie others on the topic of molecular clocks and I still struggle to find a clear and definitive answer. If a GOI can place 170 vertebrate species in their correct hierarchical orders does that not qualify it as a clock gene?

At the moment I'm using the following parameters on a 213 species data set:

lset nst=6 rates=invgamma ngammacat=4;

From my understanding this is a GTR substitution model and one that is as good as any for nucleotide analysis. I could try a codon model but I fear that the computational cost is going to be prohibitive on a large data set.

Mind you, when I used a codon substitution model on a 35 species data set not only did the Naegleria behave exemplary, but the invertebrates did too (see attached PDF). For that particular analysis I used the following parameters:

lset nucmodel=codon omegavar=ny98;

report possel=yes;

report siteomega=yes;

I ran the analysis for 12 million generations until the SD of split frequencies was less than 0.01. 

I may have to consider using the codon model for the large data set and run it on a high performance computer. 

The phylogenetic analyses have revealed some interesting phenomena. For example, rodents divide into two very distinct clades, a muroidae/dipodidae (myodonta "low rodents") clade and a sciuridae/hystricomorpha ("high rodents") clade. When running ML analyses in MEGA the myodonta clade leaves the Euarchontoglires and go and join the Afrotheria. The conclusion I draw from this is that myodonta as an sub-order have invested less in DNA proof editing than the other higher rodents (and primates) and that results in proteins that appear to be "older", hence this clade ranking with the Afrotheria in the ML analysis. 

I've run the coding sequences through HyPhy on MEGA6 and the myodonta stand out as having significantly higher codon substitution rates compared to both primates and higher rodents. I've yet to get a clear fix on the dN/dS ratio of this data and have tried to make sense out of the Kr/Kc feature in the coevol but that program comes with little in the way of instructions. Interestingly, when I run analyses using mammalian CDSs  both rodent clades join up with a common ancestor.

Anyway, the whole point with these phylogenetic analyses is not to resolve the chordate phylogenetic hierarchy but rather to lend support to the assertion that my GOI is essential to chordate/vertebrate development. BTW Artur, there's a spot available on the authors list for an intellectual contribution to the phylogenetic analysis. 

Just a quick tip considering time constraints. If your problem is convergence and you need it done really quickly, you might consider using ExaBayes instead of MrBayes.

http://sco.h-its.org/exelixis/web/software/exabayes/

The input files and the way the analyses is set up are very similar between the programs, and ExaBayes is MUCH faster due to its improved algorithms and easy parallelisation. Cheers, Fabricius

I'll check out Exabayes--thanks for the link Fabricius.  I'll see if get the IT crowd downtown to set it up on their high performance computers. As it happens the guy in charge is good colleague.

I've set up two codon model analyses on MrBayes on my desktop. I aligned the sequences in MEGA using both Clustal and Muscle and running the following parameters in MrBayes:

begin mrbayes;

set autoclose=no;

lset nucmodel=codon omegavar=ny98;

report possel=yes;

report siteomega=yes;

mcmc ngen=1000000 printfreq=100 samplefreq=100 nchains=4 nruns=2 savebrlens=yes diagnfreq=1000 stoprule=yes stopval=0.0099;

sump relburnin=yes burnin=0.25;

sumt relburnin=yes burnin=0.25 contype=halfcompat showtreeprobs=yes;

end;

You're right about the codon model being slow Artur--both runs still got 377 hr to go so I'll definitely have them run on the HPC come Monday. Sixteen days for a mere 1 million generations; if I'm going to wait a couple of weeks for results then I may as well run the analyses for at least 50 million generations or so.

But even after a mere 38,000 generations the tree topologies are looking quite decent. I take intermittent peeks at the trees by copying the *.t files and making trees using the TreeAnnotator program in BEAST1.8.  

And while the codon models are chugging along, I've got a GTR nucleotide analysis running on the same data set, but this after I tidied up the Naegleria sequence. I edited out rubbish sequence both fore and aft that had zero homology with metazoan species, but which still left me with over 80% of the coding sequence. That should reduce the LBA artefact considerably. The preliminary trees at 3 million generations are looking promising.

I would simply make a protein tree for the deep relationships and a nucleotide tree to fill our the details of more recent relationships within each of the major clades from the protein tree.

Thanks for the suggestion Patrice. I have considered that option. One of the more intriguing findings with these phylogenetic analyses is that the rodents separate into two distinct clades that don't have a direct common ancestor, especially when analysing protein sequences. The separation of these clades are anything but random; the analyses can distinguish between the muroidea (mouse, rat, hamster, Galilee blind mole rat, prairie vole, prairie deer mouse, Egyptian jerboa) and the Hystricognathi/Sciuromorpha (degu, Guinea pig, chinchilla, African mole rats, squirrels and gophers). I've analysed mammalian coding sequences using either GTR or HKY85 nucleotide substitution models and with maximum likelihood on MEGA6 the two clades are reunited with a common ancestor but still maintain their separate clades. HyPhy codon analyses reveals that the muroidea clade has a significantly higher dS rate compared to Hystricognathi/Sciuromorpha, whereas the dN rate is not significantly different between the two. Which begs the question why, if the protein sequences are not significantly different between the two clades, does the muroidea appear to look "older" than what they actually are. So much so that when I've run ML analyses of the protein sequences the muroidea rank together with the Afrotheria. I suspect that there are a handfull of key residues that are changed in the muroidea which are enough to kick them out of the Euarchontoglires. When I use the JTT protein model in MrBayes at least the mudoidea stay within the Euarchontoglires.

I know that you have a lot of invertebrates and other organisms that are very far from mammals. This can make finding a proper out group difficult and finding a model of sequence evolution that fits the mammals as well as some invertebrates should be nearly impossible. So you have a high likelihood of getting any thing but spurious results for some parts of the tree.

How many total taxa do you have? How comprehensively are the mammals sampled? And how long are your sequences? You are running an analysis on a very, very broad set of organisms that are all very distantly related. This could make even the very reliable ML method fail as likewise MrBayes for many reasons other than the latter argument can give very poor and misleading results.

As a mammalogist, I can say with some confidence that the muroids have always been known to quite distinct from the hystricomorphs and Sciurimorphs. But plenty of molecular evidence in the past 20 years confirms that Rodents are indeed monophyletic (even though for a short time before nucleotide models were improved, they were proclaimed to be polyphyletic as you have found with your data set). 

Afrotherians are such a motley, odd group of distantly related members (very long branches) since their common ancestor that they may simply be grouped together because none of them are closely related to any other mammals that they end up in a wastebasket. And so any other mammals like the hedgehog or peculiar rodents can end up in that group simply because of the best models may be inadequate to address the bias in the sequence. Then they are added to the oddball Afrotherian group in error.

Finally you should know that a gene that shows evidence of significant selective pressures is naturally prone to convergence and so is not a good gene for discovering phylogenetic relationships. The selection completely biases evolutionary history. So finding that some rodents pair with the Afrotherians raises some nice questions about the source of the bias - the reason that the sequences falsely attract one another in a phylogenetic analysis. 

An extreme example would be to use lysozyme genes to build a phylogeny which is guaranteed to be misleading because the genes arose independently at least 4 or 5 times.