Re: New paper on Neoaves

[Tim Williams <twilliams_alpha@hotmail.com>]

David Marjanovic wrote:

> Why? Of course you can. More data can never hurt, and as long as you 
> haven't reached a certain (high) limit of characters per taxon, it actually 
> adds signal. The idea is that the signal adds up while the noise cancels 
> itself out.

Yes, so runs the theory.  As Huelsenbecket al. (1996) put it: "... it is 
often argued that different data (e.g. genes) that are evolving at different 
rates may interact positively to resolve different levels of a phylogenetic 
tree. A slowly evolving gene might be useful in resolving older evolutionary 
splits but be of little use for younger groups, whereas rapidly evolving 
genes might be best for accurately resolving recent speciation events."  He 
goes onto examine whether this is always actually true; and how would we 
know if it were not true?

Huelsenbeck, J.P., Bull, J.J. and Cunningham, C.W. (1996). Combining data in 
phylogenetic analysis. TREE 11: 152-158.

I agree that more data probably can't hurt (although see below), but it 
isn't a panacea for a flawed phylogenetic method.  This 
more-data-is-always-better can be seen as an application of the Law of Large 
Numbers: as more independent trials are conducted, there will be a tendency 
for these trials to converge on the correct answer with greater frequency.  
Thus, as more data are collected, phylogenetic noise will cancel out, and 
the true phylogenetic signal will come through.  By 'noise' I mean homoplasy 
(i.e., long-branch attraction) - and I figure you do too.  However, you have 
to take into account functional constraints.  This is especially important 
for genes that encode proteins, given that individual nucleotide sites are 
not equivalent in their ability to change, so they can't be regarded as 
independent characters.  On top of this we have codon biases, especially 
important for six-fold redundancy.  I know you address this, but this isn't 
the whole story.  Protein sequence is constrained by function: there's only 
so much wiggle-room at each amino acid site before the protein becomes 
dysfunctional (or, rarely, takes on a whole new function).  As Naylor and 
Brown (1998) put it:

"...once a sequence sample of reasonable size has been obtained, accurate 
phylogenetic estimation may be better served by incorporating knowledge of 
molecular structures and processes into inference models and by seeking 
additional higher order characters embedded in those sequences, than by 
gathering ever larger sequence samples from the same organisms in the hope 
that the historical signal will eventually prevail."

Naylor, G.J.P. and Brown, W.M. (1998). Amphioxus mitochondrial DNA, chordate 
phylogeny, and the limits of inference based on comparison of sequences. 
Syst. Biol. 47: 61-76.

> Why should the noise conspire to create a false signal?

Back to long-branch attraction.  Under certain conditions, the incorrect 
phylogeny may not only be recovered, but it may actually become more robust 
as more data is added!  This is covered by Huelsenbeck et al. (1996).

> Clearly, the new Neoaves analysis is not ideal. But what are 5007 bp? I'm 
> currently writing about one that used the entire mitochondrial genome --  
> 11069 bp, of which some 6000 are parsimony-
> informative IIRC -- and still produces a suspect clade whose existence 
> should be tested by better taxon sampling.

Ah yes, mitochondrial genomes.  Check out Naylor and Brown (1998) on that 
one.  Also, have you seen Miya's work on reconstructing teleost phylogeny 
with whole mitochondrial genomes?  Some of the new mitochondrial-based 
teleost clades are just plain weird.  And at least with teleosts we have a 
highly speciose group - no shortage of taxa to sample here!  Reminds me of 
birds.  However, what about groups with only a relatively small total number 
of extant species - like crocs, or ratites, or perissodactyls, to name a 
(very) few.  Here we really hit the wall when it comes to taxon sampling.

> If there is a signal, it will add up. The noise will cancel itself out. The 
> only way that I can think of to get a fake signal is by having a skewed 
> base frequency distribution, and outside of parsimony, this is among the 
> first things analyses take into account these days.

What kind of skewed base frequencies are you referring to?  For example, 
base compositional bias is a potential problem, and the analysis may pick up 
the 'fake' signal instead of the 'true' signal.  If pervasive enough, such 
biases will be amplified with increasing data.  What began as 'noise' can be 
magnified into a robust but erroneous topology.

Cheers

Tim