Dear Michael,
the purpose of our paper (http://www.sciencemag.org/content/336/6084/1030) was to develop and document methods that enable to check for yourself what the best high resolution cutoff is - no rules of thumb are necessary!
However, repeated use of the "paired refinement technique" (3 experimental, and one synthetic datasets are documented in the paper, but evidently I've used it more often) has convinced me that values of CC1/2 between 0.1 and 0.2 in the highest shell make refinements yield better models. The difficulty here is to do the paired refinement correctly: you _cannot_ (_shouldn't_?? oh I sometimes wish I were a native speaker) compare Rfree (or Rwork, or Rwork-Rfree) for _different_ sets of reflections, which means that in order to get a meaningful information you have to compare at the _lower_ of two high-resolution cutoffs between you which to decide. And once you've identified the better cutoff, you repeat this procedure at an even higher resolution - until you find a cutoff where inclusion of the data into refinement does not improve the model any more. This is what we mean with "paired refinement".
Back to the rules of thumb: there is a strong (I find) argument for a value of CC1/2 which results in CC* of 0.5. That value is around 0.15. Please see the the discussion in the Appendix of the Henderson EM paper which we cite.
In case of experimental phasing, a CC1/2 (called CC_anom in this case) of 0.3 has been find useful for substructure solution. And data down to this value are used without maximum likelihood methods, so a lower value for refinement makes sense.
So please confirm this for yourself, and report here and/or in your paper!
hth,
Kay
On Sun, 10 Jun 2012 06:41:46 -0400, Michael Roberts <[log in to unmask]> wrote:
>Dear Crystallographers,
There have been comprehensive discussions on the use of criteria such
as Rmerge and I/sigma(I) as criteria to define the resolution to which
X-ray data is to be used for structure determination. However, CC 1/2,
the correlation between random half sets of data, is increasingly being
considered as a better means for evaluating the resolution cutoff
(Karplus and Diederichs). Is there a resolution cutoff that can be
considered as a acceptable standard, say CC 1/2 = 0.5, or does this
prompt the counter-question 'how long is a piece of string?' in which
case resolution values are to be quoted at CC 1/2 values of 0.2, 0.3,
0.5, 1, etc?
Best wishes,
Michael
Michael Roberts
Crysalin Limited
Cherwell Innovation Centre
77 Heyford Park
Upper Hey ford
Oxfordshire OX25 5HD
|