To support James Holton, MAD Scientist
From the very far past:
In 60's we have estimated X-ray diffraction data by eyes, using strips of
intensities made with propagating exposure factor increase of 2 (I guess in F
it corresponds to sqrt(2)~=1.4).
We have used to estimate our Weissenberg diffraction data 3 times (we used
X-ray films). After 1 time, logbooks were locked in the deposit safe box by our
advisers. The second round of estimation we recorded in a different logbook,
and it was again locked in the safe deposit box. After 3 time we have punched
(binary code) our data on cards, translated to ASCII, checked, corrected, and
finally run averaging programs. Our best data were of about 30% Rmerge.
However, as for the structure refinement (and maybe some flavours of structure
determination) absolute error in a measurement of a single set of symmetry
related reflections (precision) is much less important than relative error in
measuring of other sets (accuracy), we were able to refine anisotropically
small molecule structures to R=5%, because our data were not precise, but
accurate. BTW some erroneous 'discoveries' in metalo-organic complexes with
non-centrosymmetric space groups were made, based on our inability to measure
anomalous signal (it is in the single set of symmetry related reflections),
differences were averaged and absolute structure information was lost
introducing artificial asymmetry of the coordination sphere. About 20 year
later in H. D. Flack (1983). "On Enantiomorph-Polarity Estimation", Acta Cryst
A39: 876–881, all become clearly explained.
To conclude :
I am not impressed by very low Rmerge. Once ALL
reflections were overexposed on XENTRONIX (which was the area detector with not
very good dynamic range), as a result of running after low Rmerge without
understanding procedures and instruments to measure diffraction, and Rmerge of
1% was seen, but data were useless.
I am not depressed by diffraction data of very high
(but correct) Rmerge of 13% with which a structure of important for us protein
complex was swiftly solved by SAD (450 residues, 5 Se). The crystals were
small; even 10 sec exposure produced relatively weak (but accurate) data. And
long live a bending magnet that almost never burn down cryo-maintained crystals!
All depends on circumstances
My 2 cents...
Dr Felix Frolow
Professor of Structural Biology and Biotechnology,
Department of Molecular Microbiology and Biotechnology
Tel Aviv University 69978, Israel
Acta Crystallographica F, co-editor
e-mail: [email protected]
Tel: ++972-3640-8723
Fax: ++972-3640-9407
Cellular: 0547 459 608
On Mar 29, 2013, at 20:52 , James Holton <[email protected]> wrote:
> Ahh. But what I'm saying is that Rmeas is not a replacement for
> Rmerge because Rmeas is _always_ lower than Rmerge. It is even less
> useful that a low-multiplicity Rmerge for evaluating the
> diffractometer.
>
> I fully realize that Rmeas does have the desirable property of being
> "flatter" with respect to multiplicity, but being equally "too low"
> for all multiplicity is not better than being "too low" for some
> multiplicities. IMHO. Yes, I know, we all like R statistics that are
> lower. Indeed, by using the mean absolute deviation |I-<I>|, Uli was
> able to come up with a definition of Rmerge that will always be lower
> than the RMS error (for infinite multiplicity and RMS 5% error you
> actually get Rmerge=3.99%). No doubt, this must have contributed to
> the acceptance of Rmerge at the time. But we can't just keep
> re-defining our metric of "error" every ~20 years so that the same
> crappy data keeps looking better and better. That's a slippery slope
> I'd rather not be on. I think it is important to remember what it is
> we are trying to measure, and to be honest and consistent about what
> the error bars really are.
>
> But that's just my opinion. I could be wrong.
>
> -James Holton
> MAD Scientist
>
> On Fri, Mar 29, 2013 at 10:28 AM, Tim Gruene <[email protected]>
> wrote:
>> -----BEGIN PGP SIGNED MESSAGE-----
>> Hash: SHA1
>>
>> Hi James,
>>
>> you misquote me: I was saying that Rmeas should be replacing Rmerge,
>> and I guess everything you say holds for Rmeas, too, but still it is a
>> better statistical quantity than Rmerge. So please replace Rmerge with
>> Rmeas.
>>
>> Best,
>> Tim
>>
>> On 03/29/2013 06:08 PM, James Holton wrote:
>>> I must disagree with Tim on the statement "Rmerge should not be
>>> published anymore". That would be a shame. Perhaps even a crime.
>>>
>>> When Uli Arndt introduced Rmerge he was in no way, shape or form
>>> proposing that it be used for resolution cutoffs, or anything else
>>> about the quality of the "structure". He was, however, trying to
>>> define a good statistic to evaluate a diffractometer system, and
>>> Rmerge is still VERY useful for that!
>>>
>>> Any halfway decent modern detector/shutter/beam system should be
>>> able to measure reasonably strong spots to within 5% of their
>>> "true" intensity. Note that this is the _overall_ Rmerge value.
>>> The Rmerge divided up in resolution bins is pretty useless for
>>> this, especially the outermost bin, where you are basically
>>> dividing by zero. The only useful Rmerge "bin" is actually the
>>> lowest-angle one, where the spots tend to all be "strong".
>>> Remember, Rmerge is defined as the _sum_ of all the variations in
>>> spot intensity divided by the _sum_ of all the intensity. This
>>> should never be much more than 5% for strong spots. If it is, then
>>> something is wrong with either your detector, or your shutter, or
>>> perhaps your assumptions about symmetry.
>>>
>>> Yes, I know multiplicity makes Rmerge higher, but in actual fact
>>> multiplicity makes Rmerge more "honest". It is better to say that
>>> low multiplicity makes your Rmerge appear too low. Basically, if
>>> you actually do have RMS 5% error per spot, and you only measure
>>> each hkl twice, then you expect to see Rmerge=2.8%, even though the
>>> actual error is 5%. And of course, if you measure 1e6 photons in
>>> one spot you might fool yourself into thinking the error is only
>>> 0.1%. Its not. On the other hand, if all your spots are weak,
>>> then you do expect the variation to be dominated by photon-counting
>>> error, and you will get Rmerge values much greater than 5% on a
>>> perfectly good detector. It is only at high multiplicities with
>>> strong spots that Rmerge truly shows you how bad your equipment is.
>>> This is why its always good to check Rmerge in your lowest-angle
>>> bin.
>>>
>>> Yes, I know we probably all take our local well-maintained and
>>> finely-tuned beamline for granted, but that does not mean we
>>> should stop using the only statistic that tells us something might
>>> be wrong with the machine we used to measure our data. That is
>>> definitely worth the ~20 extra bytes it takes up in your paper.
>>>
>>> -James Holton MAD Scientist
>>>
>>> On Fri, Mar 29, 2013 at 6:48 AM, Tim Gruene
>>> <[email protected]> wrote: Dear Hamid,
>>>
>>> the statistics for I/sigI and the R-value per resolution shell
>>> would shed more light than the overall values.
>>>
>>> Judging from the Rmerge in the high resolution shell the data may
>>> have been processed by somebody who still thinks Rmerge <= 30% is a
>>> good criterium for resolution cut-off.
>>>
>>> The high overall Rmerge might indicate a wrong space-group was
>>> picked with too high symmetry.
>>>
>>> If you have a copy of the unmerged data, run it through pointless,
>>> if you even have a copy of the frames, reprocess them in P1 and run
>>> the data through pointless!
>>>
>>> If these data are from an article you are refereeing please point
>>> out that Rmerge should not be published anymore and be replaced by
>>> Rmeas (alias Rrim)!
>>>
>>> Best, Tim Gruene
>>>
>>> On 03/29/2013 02:19 PM, hamid khan wrote:
>>>>>> Dear CCP4BB Members,
>>>>>>
>>>>>>
>>>>>>
>>>>>> I am interested in your expert comments/opinions about two
>>>>>> values of a protein crystal diffraction data. Basically I am
>>>>>> new to this field and do not have much idea about diffraction
>>>>>> data interpretation and crystallography software’s use.
>>>>>>
>>>>>>
>>>>>>
>>>>>> 1) What could be the possible reasons for a high Rmerge
>>>>>> value, say like 0.185?
>>>>>>
>>>>>>
>>>>>>
>>>>>> 2) Value 6.2 for average I/sigma(I) for higher shell means
>>>>>> that the resolution of the diffraction data is much higher
>>>>>> than actually measured, what could be the possible reasons
>>>>>> for this?
>>>>>>
>>>>>>
>>>>>>
>>>>>> For your ease I would like to past the table here;
>>>>>>
>>>>>>
>>>>>>
>>>>>> Values in parentheses are for the last resolution shell
>>>>>>
>>>>>> Space group P2221
>>>>>>
>>>>>> Unit-cell parameters (A°)
>>>>>>
>>>>>> a 58.08
>>>>>>
>>>>>> b 101.32
>>>>>>
>>>>>> c 103.47
>>>>>>
>>>>>> Molecules in ASU 1
>>>>>>
>>>>>> Resolution range 38.63 - 2.50 (2.59 - 2.50)
>>>>>>
>>>>>> Total number of reflections 228902
>>>>>>
>>>>>> Number of unique reflections 21600
>>>>>>
>>>>>> Completeness (%) 99.1 (98.0)
>>>>>>
>>>>>> Rmerge 0.185
>>>>>> (0.373)
>>>>>>
>>>>>> Reduced χ2 0.94 (1.01)
>>>>>>
>>>>>> Average I/σ(I) 9.8 (6.2)
>>>>>>
>>>>>>
>>>>>>
>>>>>> Thanks for the tips..,
>>>>>>
>>>>>>
>>>>>> Hamid Khan
>>>
>>>
>>
>> - --
>> Dr Tim Gruene
>> Institut fuer anorganische Chemie
>> Tammannstr. 4
>> D-37077 Goettingen
>>
>> GPG Key ID = A46BEE1A
>> -----BEGIN PGP SIGNATURE-----
>> Version: GnuPG v1.4.12 (GNU/Linux)
>> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>>
>> iD8DBQFRVc8mUxlJ7aRr7hoRAg8aAKDMxBrbHy0chAnnZasHZyQHjwfysgCgu+dv
>> wEgawnqMAPQQh7wjgUv+27Y=
>> =1veF
>> -----END PGP SIGNATURE-----
