*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
It's true that thin shells don't completely solve the problem of
cross-talk, but they are much safer than random selection, in which the
NCS-related positions may well be completely surrounded by working-set
reflections and can even be arbitrarily close to one.
Besides ML issues, thick shells can make it difficult to compute Rfree
as a function of resolution, since many bins will no longer contain any
test set reflections. This can even happen with thin shells. And while
the use of NCS-related islands is another option, as pointed out, it
gets real messy with high NCS and you've still got edge effects. There
may not be an ideal way to choose a test set in the presence of NCS, but
any option is probably better than random selection. Overall, thin
shells strike me as the best compromise currently available.
It's also a good point that NCS can help improve the quality of
structures in a lot of other ways. If properly accounted for, its
benefits clearly outweigh any drawbacks.
Dean
Bart Hazes wrote:
*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
Yes SFTOOLS can select reflections in thin shells but I must admit I'm
not overly enthusiastic about this approach myself and would only use it
for high symmetry NCS cases. The problem is that to get a reasonable
number of shells covering the full resolution range, they need to be
very thin. This puts nearly every RFREE reflection at the edge of the
shell. The NCS-related reciprocal space position of an RFREE reflection
at the edge of the shell is likely going to be correlated to non-RFREE
reflections just outside the shell. I believe one person (Ian Tickle???)
suggested to use just 3 or so thick shells which is the more purist
approach but I'm not sure refinement programs can derive proper Sigmaa
weights for ML refinement when many resolution bins don't include any
RFREE reflections.
I've started long ago on an algorithm to select Rfree islands but never
completed it. Basically, you select your first Rfree reflection at
random and then flag reflections in a small NCS-related reciprocal
volume as a "no-take zone" for future Rfree reflection picks. With more
complex NCS relationships and the complexities of asymmetric unit
definitions this gets real messy real quick. However, now that NCS is
becoming so common maybe I should revisit that idea.
Bart
PS: Although the Rfree-bias problem is real, high NCS itself is also a
powerful means to reduce overfitting with properly applied restraints
and by NCS averaging of maps prior to visual inspection. So high NCS
represents both a problem and a solution.
Dean Madden wrote:
*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
Actually, one can use SFTOOLS, which is part of CCP4. However, the
selection of a test set can only be done using the stand-alone version
of the program. The "RFREE frac SHELL" command is not currently
available through ccp4i.
Dean
Jan Abendroth wrote:
A way to avoid biasing Rfree values is to choose the test set in
thin resolution shells whenever NCS is present. Currently, this
precaution is often ignored. It should become a de facto standard
for publication of structures containing NCS.
Hi all,
btw - it would be fantastic and certainly would encourage us to use
it more often if the assignment of free reflections in resolution
shells was incorporated in a ccp4 program. Yes, one can for instance
go through shelxpro, however if one wants to go back to ccp4 this
route is a bit painful.
Cheers
Jan
--
Dean R. Madden, Ph.D.
Department of Biochemistry
Dartmouth Medical School
7200 Vail Building
Hanover, NH 03755-3844 USA
tel: +1 (603) 650-1164
fax: +1 (603) 650-1128
e-mail: [EMAIL PROTECTED]
