Text 447, 66 rader
Skriven 2004-10-21 09:48:00 av Guy Hoelzer (1:278/230)
Ärende: Re: Can You Please Help M
=================================


in article cl65sm$atf$1@darwin.ediacara.org, Michael Ragland at
ragland66@webtv.net wrote on 10/20/04 10:05 AM:

> You write, "Use of computer simulations in this
> regard should be treated cautiously. It is very difficult to avoid
> assuming some (if not all) of the same aspects of the evolutionary
> process in the generation of the data and in the likelihood model
> assumed in the analysis. This results in a circularity that can easily
> mislead inferences in favor of the maximum likelihood method. There is
> no question that ML can be misled by LBA, like any other topology
> estimation method constrained to yield pictures of bifurcating trees if
> the assumed evolutionary model is incorrect, which is ALWAYS true in
> practice." Am I to take it from this last statement you consider using
> computer simulation to generate sequence data for protein families, and
> studying the influence of among site rate variation on the performance
> of phylogenetic reconstruction as useless?

No.  I am actually a big fan of modeling evolution with computer
simulations.  However, I think the the way that it has been used to compare
the efficacy of alternative phylogenetic estimation methods has often biased
the results in favor of methods that assume relatively detailed evolutionary
models.  When you simulate evolution to generate the test data sets, you
must assume a particular model of evolution.  The analytical methods also
assume various evolutionary models.  It shouldn't be surprising that "the
winner" is generally the method which assumes the same (or most similar)
evolutionary model as the one under which the data were generated.  Because
methods like maximum parsimony or neighbor joining assume generic and vague
evolutionary models, the models assumed for the generation of data usually
don't resemble those models very well.  On the other hand, the models
assumed by maximum likelihood or Bayesian methods are designed to reflect
the particular way we think evolution works in some detail, so those are
very much like the ones assumed in the generation of the data.  I don't see
this approach as a fair playing field, so I take the resulting inferences
with a big grain or two of salt.  I far prefer to use real biological data
for such comparisons.  Of course, the problem there is that we usually don't
know the Truth about the underlying phylogenetic tree, but there are
exceptions.  A classic example is the experimental evolution work by Hillis
and colleagues on the virus T7.  These authors created an evolutionary tree
in the lab with this fast evolving virus, so we can check the data analysis
against the Truth without having imposed our own model of the evolutionary
process in the generation of the data.

> It states, "Three different tree topologies were used to explore the
> effect of long branch attraction with data that incorporated
> substitution bias and different degrees of among site rate variation."
> Does the "substitution bias" constitute assuming some (if not all) of
> the same aspects of the evolutionary process in the generation of the
> data and in the likelihood model assumed in the analysis?

Particular patterns of substitution bias are often assumed in maximum
likelihood analyses of phylogenetic data.  The circularity that concerns me
is not equally problematical in all such studies, so you really have to read
each one very carefully to assess the extent of the problem.

Cheers,

Guy
---
ū RIMEGate(tm)/RGXPost V1.14 at BBSWORLD * Info@bbsworld.com

---
 * RIMEGate(tm)V10.2á˙* RelayNet(tm) NNTP Gateway * MoonDog BBS
 * RgateImp.MoonDog.BBS at 10/21/04 9:48:54 AM
 * Origin: MoonDog BBS, Brooklyn,NY, 718 692-2498, 1:278/230 (1:278/230)