In article <3q6miv$gjk at agate.berkeley.edu>, kbroman at stat.Berkeley.EDU (Karl W Broman) writes:
> Regarding Dayhoff...this model is assuming that substitutions
> occur independently across sites according to a *reversible* Markov
> chain. (By reversible, I mean that the probability that an i changes
> to a j in time t is the same as the probability that a j changing to
> an i in time t. This reversibility is very much suspect. (For
> instance, it is hard to imagine what the reversibility assumption
> means physically.)
>> It is my opinion, though I have only studied this problem
> briefly, that there is more careful work being done on this problem
> with DNA sequences (see the work by Gojobori, et al.), a key reason
> being that with DNA we are estimating only a 4x4 transition matrix,
> rather than a 20x20 one. And yet the estimators given by Gojobori's
> method, I believe, are *inconsistent*, in that even with a infinitely
> large amount of data, the estimates can be off appreciably.
>> karl
> --------------------------------------------------------------------
> Karl W. Broman University of California
> Department of Statistics
> (510) 658-2544 367 Evans Hall # 3860
> (510) 642-7892 FAX Berkeley, CA 94720-3860
>kbroman at stat.berkeley.edu USA
> --------------------------------------------------------------------
I am probably out of my depth here but two comments anyway.
1) why is reversibility suspect here (biologically)??? I would have guessed
that the process could just as validly be treated as such (for almost
all cases) with real proteins. For individual cases where natural
selection "drives" a substitution in one direction it will not hold
but for neutral substitutions it will be perfectly valid (or am I being
silly?). By the looks of things, neutral changes are VERY common and
possibly the great majority of changes are such.
2) in the underlying Dayhoff model (i.e. the set of probabilities that the
model uses) the substitution i -> j IS treated seperately from the change
j -> i ..... the 20 x 20 matrix of probabilities for change over a dstance
of 1 pam is NOT symmetric. The weight matrices that folks use for
homology searches and alignments ARE symmetric but these are just
by products of the model.
Sorry if I am talking through my hat!!
Des Higgins
>> In article <1995May26.165546.38556 at ac.dal.ca> you write:
>>I'm interested in improving empirically-based amino acid
>>substitution model for use in phylogenetic analysis.
>>>>One method which may improve upon existing models (like
>>the Dayhoff models which programs like PROTDIST and
>>PROTML use) is to determine whether the propensity
>>of an amino acid (say X) to change to another amino acid
>>(Y) depends upon the structural context in which it is
>>found.
>> [lots of stuff deleted]
>> ...
