[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Subject Index][Author Index]

RE: Morpho v molecular (was Re: Tinamous: living dinosaurs)

Tim Williams wrote-

> > I don't see anything different here than what happens in morphological 
> > analyses.
> Nor do I. I guess that's my point. Molecular-based analyses are
> dogged by the same problems as morphology-based problems.
> The major difference with morphology-based analyses is that we can
> literally see what the homoplastic characters are. For example, if
> hupehsuchians and ichthyosaurs are being linked by characters that
> relate to an aquatic mode of life, and these characters evolved
> independently, then we can readily identify what those characters are.
> By contrast, all molecular characters are 'hidden'. Out of sight,
> out of mind.

I disagree.  All of the "biases" in molecular analyses can be identified 
comparatively easily too.  Transition/transversion ratios can be changed and we 
can check if clades are still supported.  Codon and nucleotide bias can both be 
measured directly.  I say "comparatively easy" because it's not like 
morphological characters' function (especially in extinct taxa) is usually 
known, and things like exaptation and different degrees of adaptation make 
interpretations even more difficult.

> > Adding genes is better even if it causes more homoplasy, because all that 
> > means is that the topology isn't as well supported as we thought,
> Not necessarily. Random noise will be reduced by adding genes, but
> non-random noise (bias) will increase. If the latter is causing
> homoplasy, then adding more genes is not going to solve the problem,
> but make it worse. You'll get a stronger signal based on homoplastic
> characters - so the wrong topology will get *higher* support.
> To use the third codon position as an example, if this position is
> saturated (or effectively so) then it has ceased to provide any
> phylogenetic signal. If this position is actually serving as a source
> of bias in the dataset, because of (say) differential substitution,
> then this will be interpreted as structure, and mistaken for the
> phylogenetic signal.

I still don't see that you've provided a mechanism for bias to necessarily 
increase with more genes while phylogenetic signal decreases.  Sure that MIGHT 
happen, but it also might be true that your original 3 gene analysis has more 
bias per true signal than a subsequent 10 gene analysis.  We simply don't know 
and can't know without the true phylogeny.  And since we don't know if the 
original small gene sample has truer data in it than any other sample, it's 
best to take as large a sample as possible to ensure at least the topology 
isn't an artifact of the genes that heppen to have been analyzed.  You might as 
well argue our larger modern morphological analyses are making good ol' 
simplistic Romerian classifications worse because by taking more characters 
into account, they're adding homoplasy, some of which must be bias.  

> > and the "non-random noise" may be a signal of the true phylogeny.
> Yes, exactly. But how do we know the difference? The algorithm is
> searching for structure. Both non-random-noise and the phylogenetic
> signal generate structure. Further, both can be manifested as bias.

We don't know the difference, so we have to take it all and hope the majority 
of the data as analyzed by our programs represents the true tree, just like we 
have to for morphological analyses.

Mickey Mortimer