[Taxacom] Why character-tracking doesn't happen?

[Pierre Deleporte]

tentative nibbling...

when I use standard parsimony to infer a tree, e.g. from morphological 
or behavioral characters, my evolutionary assumption (model of character 
evolution) is that these characters evolved independently ("No Common 
Mechanism"), hence the analysis "column by column" is the sensical thing 
to do rather than implementing a model of concerted evolution through ad 
hoc procedures

I interpret the "robust" parts of the tree as a likely phylogeny (even 
if partially resolved)

on this phylogeny I optimise the characters, and I obtain optimal 
evolutionary scenarios, simple or complex (homoplastic)

up to this point, I can't identify my approach with what you are 
describing below (if not a tentative phylogeny, what is the tree, and 
for what use?...): I rather explicitly try to infer a phylogeny (= a 
sketchy historical explanation of the distribution of characters in taxa)
and to make use of it : "secondary" inferences, effectively of the form 
"if this phylogeny is true, then..."   


if I understand well your proposition, the secondary inference of 
homoplasy for some characters would make the phylogeny inacceptable

I have no general philosophical objection to this: the question is 
simply the degree of confidence (relative "weight" or "cost" in jargon) 
you would put on your evolutionary model that homoplasy is impossible, 
highly unlikely, slightly unlikely... for this or that character (please 
note that "homoplasy" = scattered distribution on a phylogenetic tree 
may come from different evolutionary processes, I do'nt develop this 
point here but it contributes to the complexity of the problem; e.g. an 
unlikely convergent character could be accepted as transferred through 
an episode of hybridization)

- suggestion for saving computing time in your approach: if you consider 
that homoplay is impossible for some characters, you better introduce a 
constraint in your analysis so that any topology  implying homoplasy for 
these characters is discarded from the beginning - they can't pop up as 
a result, your problem is limited to choosing among the remaining set of 
topologies, possibly very limited in some cases (easy job if it boils 
down to zero except a big unresolved radiation, an efficient protection 
against any possible homoplasy - another being clique analysis: retain 
only the subset of completely congruent characters and publish a nice 
CI=1...)

- limitation : of course you'll have to convince your pairs that you 
know that homoplasy is impossible for the characters in question (or 
that you can charge them with some relative "weight" or "cost" in the 
analysis), and these considerations should be explicitely specified from 
the start as an "enriched" model of evolutionary processes for phylogeny 
inference from these characters;
"enriched" or "biased" model, that is the question...

best,

Pierre



Bob Mesibov wrote:
> There were no serious nibbles at the bait I dangled here recently on
> character-tracking, i.e. investigating apparent homoplasy as a way to
> evaluate the plausibility of a phylogenetic hypothesis.
>
> I'm now aware that there are two philosophical objections to what I
> suggested. First, some cladists don't think of trees as history.
> Homoplasy is just a nuisance in their attempts to build a robust
> classification. This classification *reflects* history but isn't the
> same as history, so character-tracking would be meaningless.
>
> Second, some people accept homoplasy as an explanation for character or
> gene-history incongruence when *building* trees, but not as an
> explanation for *understanding* what happens in the tree they built.
>
> It works like this. A tree is built on an optimality criterion, say
> parsimony. Within this tree there's incongruence: not all the
> morphological characters, or all the sequences in a multi-locus gene
> set, will give the same tree when analysed separately. This incongruence
> is explained *after the fact* as due to homoplasy.
>
> The 'after-the-fact'er' stands by the phylogenetic hypothesis as simply
> the best available under the optimality criterion used in the analysis.
> It may or may not reflect historical reality. There is therefore no
> reason to track characters through the tree. My earlier, imagined, inner
> conversation - 'Hmmm. If this tree is correct, then this character must
> have had this interesting history...' - never happens. Optimality slays
> homoplasy, but there's no corpse.
>
> I think the same loss of information occurs when people build supertrees
> from separate phylogenetic hypotheses, or combine data before analysis
> using 'total evidence'. [Which is no such thing. It's short for 'data I
> can code and treat as equivalent and independent', and excludes relevant
> data that can't be coded in that manner.] Both methods give consensus
> trees in which the entities being 'treed' are nearly indefinable
> (Richard Zander says combining apples and oranges and lemons gives you
> fruit salad), and in which homoplasy is an explanation at a level deeper
> than the one at which the analysis operates.
>
> Do people working at this higher level think something like: 'The genes
> and/or morphological characters disagree. I don't know why and it's not
> feasible to investigate. All I know is that this is a pretty good tree
> based on sound methodology'??
>   

-- 
Pierre Deleporte
Université de Rennes 1
CNRS UMR 6552 
Station Biologique de Paimpont
F-35380 Paimpont   FRANCE
Téléphone : 02 99 61 81 63
Técopie : 02 99 61 81 88