Comparing stats on geometry as well as R values is confounded by the problem of choosing weights. Should you hold the weights fixed or perform an individualized weight optimization at each step?
More importantly, our geometry libraries are imperfect. We know from surveys that the fit of models to Engh & Huber gets worst as the resolution of their X-ray data gets higher. This is because there are real (and to a good extent conformationally dependent) variations in bond angles that are brought to light by the high resolution (better than about 1.6A) data. If I add quality high resolution data I would expect the geometry stats to get worst. Dale Tronrud On 7/7/2015 11:17 AM, Shane Caldwell wrote: > Chiming in late with a follow-up question: > >>On the other hand in paired refinement, if adding the data improves the > structure >>as measured by Rfree in a zone excluding the added data, then it is > hard to deny >>that that data are worth including. > > Is it correct to think that model geometry would also be valuable at > this point? If adding reflections lead to a model with more reasonable > average bond angles, reduced clashes, etc., that would indicate that the > added reflections have improved the refinement, no? The geometry stats > should also be completely independent from the crystallographic stats. > > Shane Caldwell > McGill University > > > > On Fri, Jul 3, 2015 at 2:56 AM, Kay Diederichs > <kay.diederi...@uni-konstanz.de <mailto:kay.diederi...@uni-konstanz.de>> > wrote: > > On Thu, 2 Jul 2015 13:25:07 -0400, Edward A. Berry > <ber...@upstate.edu <mailto:ber...@upstate.edu>> wrote: > > >My take on this- > >No one has been willing to specify a cutoff (and probably there is no > rigorous way to > >mathematically define the cutoff) and say "If CC* (or CCfree or > whatever) is below X > >then it will not improve your structure, if above X then it will". > > the electron microscopy community uses a similar measure ("FSC", > Fourier Shell Correlation) as CC1/2. They follow their "Gold > Standard" if they cut their data at FSC=0.143 . Mathematically, > CC1/2=0.143 is equivalent to CC*=0.5 . So researchers of that > community _do_ specify a cutoff, and by calling it "Gold Standard" > they cast it in stone. Very helpful because probably no reviewer > ever calls a "Gold Standard" into question. > > > Probably depends > >among other things on how strong the lower resolution data is, how good > the > >structure is without the added data. > > the latter point is crucial: a structure that is good can "make use > of" higher-resolution data than a structure that is not properly > refined. That is different from the situation in electron > microscopy, where the phases are obtained experimentally. This is > why an X-ray structure at an early stage of iterative > refinement/fitting may possibly not be improved by the weak > high-resolution data , and why the paired refinement should be done > during the "end game" of refinement. But as long as CC1/2 is > statistically significant it may improve a very good model even if > CC*<0.5 (see Bublitz et al IUCrJ 2, 409-420 (2015) for an example). > > To find out how close the structure to the data is, it helps to > compare CCwork and CC*. > > An arbitrary cutoff (like CC*=0.5) is thus not sensible in all > situations; it may serve as a rule of thumb though since the > difference in resolution limit is not large anyway: CC*=0.5 and > CC*=0.3 usually differ by (say) 0.1A only. > > >On the other hand in paired refinement, if adding the data improves the > structure > >as measured by Rfree in a zone excluding the added data, then it is hard > to deny > >that that data are worth including. > > Absolutely. Much better than to believe in rules of thumb. > > best, > > Kay > > > > >eab > > > >On 07/02/2015 12:52 PM, Keller, Jacob wrote: > >> Wasn’t all of this put to bed through the implementation of CC > measures? > >> > >> JPK > >> > >> *From:*CCP4 bulletin board [mailto:CCP4BB@JISCMAIL.AC.UK > <mailto:CCP4BB@JISCMAIL.AC.UK>] *On Behalf Of *Robbie Joosten > >> *Sent:* Thursday, July 02, 2015 12:46 PM > >> *To:* CCP4BB@JISCMAIL.AC.UK <mailto:CCP4BB@JISCMAIL.AC.UK> > >> *Subject:* Re: [ccp4bb] paired refinement > >> > >> But it is not the R-free of the shell here. In paired refinement > you take the R-free of the reflections outside the shell. > >> > >> Cheers, > >> Robbie > >> > >> Sent with my Windows Phone > >> > >> > > ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------ > >> > >> *Van: *Edward A. Berry <mailto:ber...@upstate.edu > <mailto:ber...@upstate.edu>> > >> *Verzonden: *2-7-2015 18:43 > >> *Aan: *CCP4BB@JISCMAIL.AC.UK <mailto:CCP4BB@JISCMAIL.AC.UK> > <mailto:CCP4BB@JISCMAIL.AC.UK <mailto:CCP4BB@JISCMAIL.AC.UK>> > >> *Onderwerp: *Re: [ccp4bb] paired refinement > >> > >> Another criterion for cutoff, also requiring the structure to be > solved, > >> is the agreement between data and structure, e.g. Rfree or CCfree. > >> I think it is very unlikely that you could get Rfree =.2493 in a > shell > >> which contains only noise. So I would suggest doing paired refinement > >> to 2.2 and 2.1 A (if the data is available). > >> > >> On 07/01/2015 07:15 PM, Eric Karg wrote: > >> > Hi all, > >> > > >> > I have a dataset processed in XDS to 2.3 A (based on CC1/2). > I'm trying to do "paired refinement" to determine the optimal > resolution cutoff. Here is what I get at different resolutions set > in Phenix: > >> > > >> > Final Rfree/Rwork: > >> > 2.7—> 0.2498/0.2027 > >> > 2.6—> 0.2519/0.2009 > >> > 2.5—> 0.2567/0.2025 > >> > 2.4 —> 0.2481/0.2042 > >> > 2.3 —> 0.2493/0.2075 > >> > > >> > The geometry of all output structures are similar. > >> > > >> > 1. What is the high resolution cutoff based on these data? I > know that Rfree/Rwork at different resolution should not be > compared, but is there a simple way to do the test as described in > the K&D 2012 Science paper using Phenix GUI? > >> > > >> > 2. For refining a structure at a lower resolution (lower than > the initial dataset), do I simply set the resolution limit in the > refinement or I need to reprocess the data starting from the images? > Do I need to do anything with Rfree flags? Based on the discussions > on this forum I know I should deposit the highest resolution dataset > but my question is about the mtz file which will be used for refinement. > >> > > >> > Thank you very much for your help! > >> > > >> > >
