On Tuesday 05 June 2007 12:19, Edward A Berry wrote:
> You have a good point there and I would be interested in hearing
> some other opinions, so I take the liberty of reposting-
>
> My instinctive preference is that each structure should be
> supported solely by the data that is deposited with it -
> (one dataset one structure) but in terms of good science
> we want to produce the best model we can, and that might be
> the rigid-body-located structure from another dataset.
I don't think that is quite the right way to look at it.
In general we refine our model so that it both
- agrees with the data
- agrees with a priori knowledge
In maximum likelihood terms: we want to find the model that is
the most likely explanation for our observed data.
An inherently unlikely model is also an inherently unlikely
explanation. Therefore we focus on likely models.
We impose geometric restraints because we believe that we have
a better a priori expectation for bond lengths and angles than
can be determined de novo from the data in this one experiment.
Similarly we impose the known sequence of our protein on the
model, even if the maps are not sufficiently good to identify
each amino acid directly from the electron density.
If we have an a priori expectation for the conformation of
the whole protein, or large pieces of it, then we should
account for this in the model, even if the data is not
sufficiently good to reproduce this expectation de novo.
Therefore if you have a high-resolution structure available,
the best treatment of low-resolution data may well be to
place the known structure as a rigid body. If you suspect
hinge motions or other large scale inter-domain shifts, you
might want to refine the hinge angle explicitly, but unfortunately
our usual refinement programs are not really set up for this.
These are important issues, and are close to the heart
of the Maximum Likelihood approach to model refinement.
Ethan
> cdekker wrote:
> > Hi,
> >
> > Your reply to the ccp4bb has confused me a bit. I am currently refining
> > a low res structure and realise that I don't know what to expect for
> > final R and Rfree - it is definitely not what most people would publish.
> > So the absolute values of R and Rfree are not telling me much, the only
> > gauge I have is that as long as both R and Rfree are decreasing I am
> > improving the model (and yes, at the moment that is only rigid body
> > refinement).
> > In your email reply you suggest that even though a refinement to
> > convergence that will lead to an increased Rfree (and lower R? - a
> > classic case of overfitting!) would be a better model than the
> > rigid-body-refined only model. This is what confuses me.
> > I can see your reasoning that starting with an atomic model to solve
> > low-res data can lead to this behaviour, but then should the solution
> > not be a modification of the starting model (maybe high B-factors?) to
> > compensate for the difference in resolution of model and data?
> >
> > Carien
> >
> > On 4 Jun 2007, at 19:38, Edward A Berry wrote:
> >
> >> Ibrahim M. Moustafa wrote:
> >>> The last question: In the same paper, for the complex structure R and
> >>> Rfree are equal (30%) is that an indication for improper refinement
> >>> in these published structure? I'd love to hear your comments on that
> >>> too.
> >> Several times I solved low resolution structures using high resolution
> >> models, and noticed that R-free increased during atomic positional
> >> refinement. This could be expected from the assertion that after
> >> refinement to convergence, the final values should not depend on
> >> the starting point: If I had started with a crude model and refined
> >> against low resolution data, Rfree would not have gone as low as the
> >> high-resolution model, so if I start with the high resolution model
> >> and refine, Rfree should worsen to the same value as the structure
> >> converges to the same point.
> >>
> >> Thinking about the main purpose of the Rfree statistic, in a very
> >> real way this tells me that the model was better before this step
> >> of refinement, and it would be better to omit the minimization step.
> >> Perhaps this is what the authors did.
> >>
> >> On the other hand it does not seem quite right submit a model that
> >> has simply been rigid-body-refined against the data- I would prefer to
> >> refine to convergence and submit the best model that can be supported
> >> by the data alone, rather than a better model which is really the model
> >> from a better dataset repositioned in the new crystal.
> >>
> >> Ed
> >
> >
> > The Institute of Cancer Research: Royal Cancer Hospital, a charitable
> > Company Limited by Guarantee, Registered in England under Company No.
> > 534147 with its Registered Office at 123 Old Brompton Road, London SW7 3RP.
> >
> > This e-mail message is confidential and for use by the addressee only.
> > If the message is received by anyone other than the addressee, please
> > return the message to the sender by replying to it and then delete the
> > message from your computer and network.
>
--
Ethan A Merritt
