RE: New paper on Neoaves

[Tim Williams <twilliams_alpha@hotmail.com>]

Michael Mortimer wrote:

> > I'd hate to think that there is a belief that you can 'improve' the 
> > phylogenetic signal simply by expanding the dataset (=stringing together 
> > more and more DNA sequences).
>
> Er, can't we?  Isn't this the whole idea behind supermatrices (not 
> supertrees)?

Except if (for example) all the genes share the same compositional biases - 
as would happen if they were pulled from the same genome (chromosomal or 
mitochondrial) from the same taxon.  Thus, by piling on more genes to 
produce a "combined dataset" you would simply be compounding the problem, by 
amplifying the bias.  If two lineages evolved this bias independently, it 
could (and apparently sometimes does) trump the 'true' historical signal.  
You'd have well-supported nodes that were completely erroneous.

> I'd think you'd get a better phylogenetic signal if you did this.  The 
> levels where divergence was poorly captured by a gene would be poorly 
> supported, and easily overwealmed by another gene that captured divergence 
> better at those levels.

Then why don't the individual genes alone capture separate levels of 
divergence?  Check out the beta-fibrinogen data w.r.t Metaves, for example.

The other problem is that, if there is a bias, it may overcome the 
individual phylogenetic signals lurking in the various genes, since the bias 
is one thing that the genes would share in common.  (Unless both 
mitochondrial and chromosomal genes are used, since each genome has its own 
repertoire of tRNA's.)

> Yes, increased taxon sampling does generally help.  For instance, an 
> analysis can have ambiguity
> due to character conflicts, but new taxa can show which of these characters 
> were convergently
> developed.  As an example, if you had troodontids in a polytomy with 
> dromaeosaurids (sickle claw)> or ornithomimids (parasphenoid bulla), adding 
> Sinovenator would resolve the tree.

Yes, because _Sinovenator_ clarified the polarity of character states at the 
base of the clade.  In the same way, basal alvarezsaurids like _Patagonykus_ 
helped resolve the position of _Mononykus_, and exposed the homoplastic 
characters that erroneously caused Mononykus_ to group inside the bird 
clade.

Unfortunately, the nature of genomes and proteins makes homoplasy a far more 
pervasive and insidious problem for gene-based analyses compared to 
morphology-based analyses.  When you get a 'bad' gene-based phylogeny (i.e., 
either poorly-supported nodes; or 'weird' clades that drastically contradict 
morphology; or both) the instinctive response is just to add more taxa, add 
more genes, or both - and stop adding taxa or genes when the phylogeny looks 
'good' (i.e., well-supported nodes; or clades that are congruent with 
morphology; or both).  Often very little attention is given to the 
underlying mechanism as to WHY this occurs, since it is assumed (or hoped) 
that this expanded gene or taxon sampling is better at capturing the 
historical signal at the expense of the 'noise'.  I don't think this is 
always the case.

> I suppose I don't get what you're worrying about.  If we have extremely 
> well (>95%) supported nodes, then we know things are working.

How I wish that were always true.  :-(

Cheers

Tim