The RFREE SHELL command in sftools is another way to select thin shells.
However, for the purist there is no clean way of getting rid of
correlations between the Rwork and Rfree sets that I know of.
To do it properly, the thickness of the thin shell should be larger than
the radius where the G-function has strong features. In practice that
makes the shells so thick that you can have only a few which likely
causes artifacts. Even then, the reflections at the edge of the shells
are still contaminated so you'd have to exclude the ones near the edge
from both the Rwork and Rfree sets. I'm pretty sure that there has been
a publication long ago on thick shells (It may have been Ian, Phil or
some other CCP4BB regular).
An improvement would be to go from shells to donuts, basically the
intersection of the spherical shell and a plane through the origin
perpendicular to the NCS rotational symmetry axis. One step further is
Fred's suggestion to use the explicit NCS relationships to select groups
of NCS-contaminated reflections, but only works if the NCS symmetry
forms a closed group. One step further again would be to use an explicit
low resolution G-function model to find the strongest NCS correlations
and only group those together.
Enough fodder for purists to fight over but when sftools users ask me,
my general response includes the fact that the higher your NCS the
greater the NCS-contamination problem, but also the smaller the model
bias/overfitting problem, assuming NCS restraints are used appropriately
in refinement. As a reviewer I would have a problem with authors patting
their backs for a small Rwork-Rfree difference without noting the impact
of NCS. But if they report a low Rwork-Rfree difference and comment on
the fact that the difference is likely underestimated due to NCS effects
I have no problems, with or without thin shells, as long as I feel the
refinement protocol and final model are acceptable for the biological
interpretations that they make.
Bart
Vellieux Frederic wrote:
> Hi Ian (& ccp4bb'ers),
>
> NCS ties reflections in reciprocal space by the interference
> G-function effect. Nothing more. So you get an R-free value that is
> lower than if you don't have NCS. One should be aware of that, and
> referees should be aware of that.
>
> I currently have a structure that has 12-fold NCS in the asymmetric
> unit. The free R-factor is lower than the R-factor. I expect that
> future referees will not view that kindly.
>
> A number of people have suggested to use different approaches to get
> rid of this reciprocal space binding effect. One of these people (Bart
> Hazes I think, correct me if I'm wrong) suggests to take reflections
> for the R-free as thin shells in reciprocal space. The thin shell is
> omitted completely from the target for refinement (I suppose omitting
> a shell of data completely will also have deleterious effects on the
> refinement, I don't know by how much). Problem is, none of the data
> processing programs or suites of programs has implemented this as far
> as I know. A better approach would be to use the NCS operator (the
> transpose and the inverse of the rotation matrices in fact, including
> the identity matrix for all cases including the cases where the only
> NCS is 1-fold NCS, i.e. the presence of solvent in the asymmetric unit
> or unit cell) to select the subset of reflections that are going to be
> omitted from the refinement target: take one reflection, select all
> equivalents that are bound to it by the interference effect, and
> repeat the process until you have reached the required number of
> reflections to be omitted. But this requires serious programming...
> And someone willing to modify all data processing suites to include
> this approach. But that would satisfy referees because it is the only
> approach that is valid.
>
> Fred.
>
> Ian Tickle wrote:
>> The problem is that real life is never simple! -
>> and NCS really messes things up!
>>
>> Cheers
>>
>> -- Ian
>>
--
===========================================================================
Bart Hazes (Associate Professor)
Dept. of Medical Microbiology & Immunology
University of Alberta
1-15 Medical Sciences Building
Edmonton, Alberta
Canada, T6G 2H7
phone: 1-780-492-0042
fax: 1-780-492-7521
===========================================================================
|