On Saturday 18 September 2010, Nicholas M Glykos wrote:
> <snip>
> > it seems that we are trying to deposit one model to satisfy two
> > different purposes - one for model validation and the other for model
> > interpretation (use in docking etc), and what's good for one purpose
> > might not be necessarily good for the other.
> </snip>
>
> This has been discussed before on this list, but allow me to repeat it:
> You would have expected that the crystallographers' aim would be to
> deposit the model that maximises the product (likelihood * prior).
> Clearly, this is not what we do,
I guess I have more faith that we do in fact aim for that.
Our data, programs, models, and insight are imperfect,
but we do our best with what we have.
> mainly because (a) the calculation of
> likelihood is only based on a subset of the 'data' that are obtained from
> an X-ray diffraction experiment (for example, we ignore diffuse scattering
> as Ian pointed-out),
I do not think that is a valid criticism. In any field of science
one might hypothesize that conducting a different kind of experiment
and fitting it in accordance with a different theory would produce
a different model. But that is only a hypothetical; it does not
invalidate the analysis of the experiment you did do based on the
data you did collect.
> (b) we consciously avoid 'prior' because this would
> make the models 'subjective', meaning that better informed people would
> deposit (for the same data) different models than the less well informed,
I don't know of anyone who consciously avoids using their prior
knowledge to inform their current work. But yes, people with more
experience may in the end deposit better models than people with
little experience. That's why it is valuable to have automated tools
like Molprobity to check a proposed model against established prior
expectations. It's also one way this bulletin board is value, because
it allows those with less experience to ask advice from those with
more experience.
> (c) the format of the PDB does not offer much room for 'creative
> interpretations' of the electron density maps [for example, you can't have
> discrete disorder on the backbone (or has this changed ?)].
Could you expand on this point?
I am not aware of any restriction on multiple backbone conformations,
now or ever. It is true that our refinement programs have not always
been very well suited to refine such a model, but that is not a fault
of the PDB format.
> I sense that
> what is being deposited is not the 'best model' in any conceivable way,
> but the model that 'best' accounts for the final 2mFo-DFc map within the
> limitations of the program used for the final refinement.
That would be true if the refinement is conducted in real space.
However, it is nearly universal to do the final refinement in
reciprocal space.
If a maximum likelihood residual is used, the aim is to achieve the
"best model" in the generally accepted formal sense of being the
the set of model parameter values that provide the most likely explanation
for the observed data. The priors are imposed as restraints;
the partial residual R_crystallographic(Fo, Fc) encompasses the agreement
with the observed data.
> My twocents,
> Nicholas
And mine in return :-)
Ethan
