Dear Kay,
I wonder what is your opinion of the following proposition.
"None of the data quality indicators derived from data alone matter too
much".
Let me explain what I mean by this.
Ultimately, I truly don't care what value of Rmerge, Rpim, or even CC1/2
data processing produces from the set of frames I toss at it. Surely it
is important to keep an eye on them to verify that dataset is kosher and
to obtain an *initial* estimate of the resolution limit of useful data.
But the actual values never deter from trying to solve the structure.
If it cannot be solved - well, then none of the aforementioned
indicators matter at all. If it is solved, the only remaining question
is how far the useful data goes. And that should be determined using
Karplus-Diederichs (KD) test. I do use CC1/2~0.5 as initial resolution
cutoff these days, but before finalizing any model I run the KD test. I
do look at where the Rpim, I/sigma and CC1/2 end up at the resolution
edge, but only out of curiosity and to adjust my perception of how they
correlate with true resolution.
And I think that efforts should be targeted at optimizing KD test as a
tool rather than being distracted by outdated approaches that were
proposed in the computationally-handicapped times.
It is entirely possible that all this is exactly what you said, just
with different wording. But I guess more wording is still needed given
that people keep asking about Rmerge.
Cheers,
Ed.
On Tue, 2013-11-19 at 14:22 +0000, Kay Diederichs wrote:
> Hi Jim,
>
> of course the issue of crystallographic data quality indicators deserves a somewhat more appropriate (or at least more permanent, and peer-reviewed) means of dissemination than CCP4BB. Nevertheless I'll sum up some of the most important points I can think of:
>
> A) all data quality indicators measure precision, not accuracy
> B) there are those data quality indicators that measure the precision of unmerged data:
> (Rsym=)Rmerge, Rmeas, (I/sigma)_unmerged
> and those that measure the precision of merged data:
> Rpim, Rsplit (the FEL community uses this; same as R_mrgd_I - see Diederichs&Karplus 1997), CC1/2, (I/sigma)_merged
> The merged indicators usually differ by a factor of sqrt(m) from their unmerged counterparts, where m is multiplicity. Rsplit (~R_mrgd_I) and CC1/2 compare random half-datasets which may be more robust than just hoping that the explicit sqrt(m) law holds (it only holds for unrelated errors). There is no unmerged counterpart of CC1/2.
> C) Since downstream steps use intensities, it is preferable to use a data quality indicator that does not require sigma to be estimated, because the authors of the different data processing programs/algorithms have different ideas how this should be done. This rules out I/sigma as a useful quality indicator - at least as soon as different programs look at the same data.
> D) Merged data quality indicators are more useful because we are using merged data for downstream steps (phasing, molecular replacement, refinement), so we need to know _their_ precision, not that of the unmerged data.
> E) Rpim and Rsplit are calculated from intensities and have a different asymptotic behaviour than model R-values (Rwork, Rfree), so they cannot be meaningfully be compared with model R-values (i.e. their numerical value tells you nothing about the Rwork/Rfree your model can be refined to). This is very different from CC1/2 - it can be used to calculate CC*, a quantity that is the upper limit of what the CC of the model intensities against the experimental intensities can reach.
>
> I'll stop here. Most of this may be at variance with what we were all brought up with, but it's time for a change!
>
> best,
>
> Kay
>
> On Tue, 19 Nov 2013 13:18:19 +0000, Jim Pflugrath <[log in to unmask]> wrote:
>
> >Graeme wrote:
> >"... Rpim is much more instructive. ... as each of these tells something different."
> >
> >I have to ask:
> >"Why is Rpim much more instructive? I'm trying to figure this out still. Can one please summarize what are best practices with all these numbers and how each of these tells something different?"
> >
> >Another problem that I see is that folks can adjust their sigmas many different ways without knowing they have adjusted their sigmas. And they can be adjusted incorrectly when they are adjusted.
> >
> >BTW, Graeme is correct about lots of multiple low I/sigI observations for each Bragg reflection in a resolution shell will lead to 100% (or higher) Rmerge with <I/sigI> of 3. This assumes no systematic errors and only randomly distributed random errors (a rare if not impossible situation, I would think). I will defer to others about what the relevance of that is.
> >
> >Thanks for any insights, Jim
> >
> >
> >________________________________
> >From: CCP4 bulletin board [[log in to unmask]] on behalf of Graeme Winter [[log in to unmask]]
> >Sent: Tuesday, November 19, 2013 2:02 AM
> >To: [log in to unmask]
> >Subject: Re: [ccp4bb] 100% Rmerge in high resolution shell
> >
> >Usually this means that you have relatively high multiplicity, which give-or-take improves the I/sig(I) by sqrt(m) where m is the multiplicity, but also increases the Rmerge.
> >
> >For any given narrow shell of reflections,
> >
> >Rmerge ~ 0.8 / unmerged(I/sig(I))
> >
> >merged(I/sig(I)) ~ sqrt(m) * unmerged(I/sig(I))
> >
> >So it is perfectly possible to have unmerged I/sig(I) of 0.8 which will give you an Rmerge of around 1.0, and have I/sig(I) (merged) around 3, by having multiplciity 14 or so. I suggest that this is the case: if it is much lower than this there is something odd going on.
> >
> >For the merged I/sig(I) Rpim is much more instructive. I'd love it if people reported merged and unmerged I/sig(I), Rmerge, Rmeas, Rpim, CC1/2, ... as each of these tells something different.
> >
> >Best wishes,
> >
> >Graeme
> >
> >Possibly useful papers:
> >
> >http://www.nature.com/nsmb/journal/v4/n4/abs/nsb0497-269.html
> >http://scripts.iucr.org/cgi-bin/paper?he0191
> >http://scripts.iucr.org/cgi-bin/paper?he0268
> >
> >
> >
> >
> >On 19 November 2013 06:43, Shanti Pal Gangwar <[log in to unmask]<mailto:[log in to unmask]>> wrote:
> >Dear All
> >
> >
> >Can anyone explain the meaning and relevance of data when the Rmerge is 100% in high resolution shell and I/sig(I) is 3.
> >
> >
> >
> >Thanks
> >
> >
> >
> >--
> >********************
> >regards
> >Shanti Pal Gangwar
> >School of Life Sciences
> >Jawaharlal Nehru University
> >New Delhi-110067
> >India
> >Email:[log in to unmask]<mailto:[log in to unmask]>
> >
> >
> >
> >
--
Edwin Pozharski, PhD, Assistant Professor
University of Maryland, Baltimore
----------------------------------------------
When the Way is forgotten duty and justice appear;
Then knowledge and wisdom are born along with hypocrisy.
When harmonious relationships dissolve then respect and devotion arise;
When a nation falls to chaos then loyalty and patriotism are born.
------------------------------ / Lao Tse /
|
