Re: More on the genus problem

[Augusto Haro <augustoharo@gmail.com>]

2009/9/21 David Marjanovic <david.marjanovic@gmx.at> wrote:

>>  Although the phylogenetic tree may be appreached with cladistic
>>  methodology, we have no certainty that parsimony prevailed in
>>  evolution.
>
> 1) That's why maximum likelihood and Bayesian analysis -- extensions of
> parsimony that take more factors into account -- were developed. Indeed,
> parsimony is more prone to long-branch attraction than they are.
> 2) However, parsimony fares better in simulation studies when the speeds of
> evolution of too many different nucleotides are different ("heterotachy").
> 3) Also, parsimony doesn't have a problem with missing data as long as a few
> informative cells remain, while likelihood and Bayesian analysis run into
> great trouble according to a simulation study in the Systematic Biology
> issue of February.

Do not know too much about this "performance" discussion, so sorry any
mistake. Until now, I made up my mind on the probabilistic
methodology/parsimony (somehow I come up equating "cladistics" with
parsimony, I think some author did it before) battlefield followed
some early work by Farris. Farris said we do not really know how
evolution operates, so it is too ad hoc for him to follow any given
evolutionary model, such as a k2p, for example, and not any other. He
said that as well as we do not know if parsimony prevailed in
evolution, neither do we know if any of these other evolutionary
models are true, and that knowledge of how evolutionary processes are
should follow previous knowledge of the phylogeny, not the other way
around. So, according to him, the only thing we can do is to see what
evidence by itself indicates. As I understand it, parsimony shows you
what in a larger measure follows from the data, because it is the only
method which reduces ad-hoceries (and I suppose if we reduce
ad-hocery, we have in a greater proportion what follows from the data
themselves). So, admitting we cannot know the true evolutionary model,
the best thing we should do is parsimony (personally speaking, this is
one of the reasons I think parsimony is better than the probabilistic
models).

> I don't see how a cladogram produced using parsimony contains more
> information than one produced using likelihood or Bayesian analysis.

When the results of Bayesian analysis or Maximum Likelihood differ in
topology from the results of parsimony, the trees produced by these
probabilistic are clearly longer under a parsimonious reconstruction
of the changes -under a not necessarily evolutionary perspective we
should use other word, which I do not remember- it implies than a MPT.
The MPT is maximally informative in that you need the smallest number
of diagnostic features for the complete group of taxa the MPT implies.
Because each node/OTU is diagnosed by its proper apomorphies, it is
clear that in an MPT the total number of diagnostic instances (would
be each apomorphy) is lesser. When you can reduce the number of
symbols necessary to transmit information (apomorphies), the
information content of these symbols is maximal taken as a whole, and
the information is more efficiently or economically organized (because
of the less symbols). Or at least that is what I came up interpreting
from these readings.

> But these didn't even exist yet in 1979.

You are right with Bayesian analysis, which seems to be no older than
the nienties, but Maximum likelihood is older than Wagner trees. And
Felsenstein came up defending these schemes from 1973 (indeed, much of
what I read from Farris are criticism to Felsenstein's ideas)

Cavalli-Sforza, L. L. and Edwards, A. W. F. (1964). Analysis of human
evolution. Proc. XI Intern. Congr. Genet., 3, 923-933.

Edwards, A. W. F. and Cavalli-Sforza, L. L. (1966). Estimation
procedures for evolutionary branching processes. Bull. Intern.
Statist. Inst., 41, 803-808.

Felsenstein, J. 1973. Maximum likelihood and minimum-steps methods for
estimating evolutionary trees from data on discrete characters. Syst.
Zool. 22: 240-249.

> I bet Farris was talking about how a
> phylogenetic tree contains more information about phylogeny than a
> classification does... which is... like... not surprising.

As far as I understood it, he referred to information content in the
sense of information theory... He was one of those guys who knew too
much about logical programming -he programmed the first parsimony
algorythms-, so he may have had some good knowledge of the theoretical
bases of what he did in practice. It would be pedantic to say that
because it is hard to me to follow his writings he may be a genius,
but from what I think I understood from him, he may be one of the most
outstanding geniuses in the field of theoretical biology of the last
decades...