In article <3gc2l3$120o at columba.udac.uu.se> gerard at rigel.bmc.uu.se (Gerard
Kleijwegt) writes:
>> As any crystallographer who uses X-PLOR knows, MD is the quickest way
> to screw up any perfectly good structure if you switch off the xray
> pseudo-energy term; a few ps at room temperature will do it.
As any person knows, using X-ray information without the theoretical
stereochemical information would result in a 'random' collection of
points.
> May I take the opportunity to suggest a simple test to see if homology
> models are better than random ? Assume the structure of protein X
> is known, and that it is > 30 % homologous to protein Y. Now, if
someone
> collects xray data of protein Y, a homology model of protein Y should
> at the very least give a clearly better solution in the Molecular
> Replacement than the structure of protein X. If not, the homology
> model does not contain more information than the structure of the
> 'undisturbed' protein X itself (and is therefore equally good as
anyone's
> guess, where the xtallographer's guess would be to use a stripped
> version of protein X).
This is a good test of a usefulness of a model, but it is certainly not
the only one.
>> - even if the fold can be predicted, this usually carries very little
> biologically-relevant information. The most interesting parts of a
> protein are often found in one or more of its loops. Moreover, at the
> end of the day, you'll only be interested in the conformation of a
> handful of side chains (the active site, a metal-binding site, a
ligand-
> binding site, a protein-protein or protein-DNA interface), something
> which even with 80 % homology is often hard to do. And nature is
always
> full of surprises. For an example, see the structure of cellular
> retinoic-acid binding protein: about 30% homologous to several other
> solved structures, known fold etc. Still, the few insertions have
> a rather drastic effect on the structure (see Structure 2(12), pp.
1241-58).
> This protein was one of the targets in the prediction competition and
> of course all predictors got the fold right, but the saillant details
> hadn't been modelled correctly by anyone as far as I know. Who would
> expect that a helix would suddenly be extended by an extra turn, that
a
Wrong. We had an extra turn in that helix in several of our models. I
talked about that at the prediction meeting in Asilomar.
> strand-turn-strand corkscrews over quite some distance, that the
> ligand does not interact with the corresponding residue with which it
> interacts in a related structure, and that the ligand actually sticks
> out of the protein rather than being "engulfed" by it ?
> Not a million years of molecular dynamics is going to model such
> subtle and intricate details (unless you have an xray or noe term
> to switch on ;-).
>>
Your view is probably a little too restrained. For example, a molecular
biologist may have a sequence and some guesses that it interacts with a
certain charged ligand. If he is interested in identifying the binding
site (and as you point out, different members of the family may bind
different ligands and in different regions), he can do ~ 50 random
mutations and hope that he will hit the site by luck. Alternatively, he
can calculate a homology model and then the electrostatic potential around
the model. Approximate locations of many binding sites can be predicted in
this way in a matter of days and so a smart and unbiased molecular
biologist, who chose to collaborate with a modeler first, can do his
mutations much more economically and faster, and so scoop another
molecular biologist who chose to collaborate with a crystallographer.
Applications like these, which only require low/medium resolution models,
are many and comparative modelling is frequently more than sufficient to
deal with them.
Let's not create a tug of war between crystallographers and modelers.
After all, we modelers depend on crystallographers for raw data and could
not do anything without them.
Regards, Andrej
