Reference for Resolution Cutoffs

From: Jacob Keller
Date: 6 December 2011 17:13


Dear Crystallographers,

I hate to broach this subject again due to its wildly controversial
nature, but I was wondering whether there was any reference which
systematically analyses resolution cutoffs as a function of I/sig,
Rmerge, Rmeas, Rpim, etc. I strongly dislike Rmerge/Rcryst for
determining cutoffs, for obvious reasons--and especially for datasets
of higher multiplicity--but nevertheless it is a ubiquitously-reported
statistic, and one therefore has to make an argument against using it.
Hopefully this could be done by pointing to a definitive reference--or
am I stuck with a convention versus the truth? Maybe the ACA or
similar could make a public anti-Rmerge proclamation about it, to make
it easier for us?

Also, more generally, it seems that the refinement programs are now
better able to discount lousy high-res data, so why not leave the
choice to those programs, and just give them all of the data to the
edge of the detector, especially since our computational and data
storage capacities are now completely sufficient for that? One could
then use some other metric for the goodness of the structure, such as
what bin crosses the Rfree = 40% mark or something.

One could push this even further and, as has been mentioned on this
list before, just give the refinement program all of the intensities
of the voxels in the 3D dataset?

Jacob

--
----------
From: Ethan Merritt

What is your question, exactly?
I don't follow the logic that because a statistic is reported, one
must therefore argue against it.
Acta published at one point a guideline as part of the instructions
to authors, but the state of the art passed it by very soon after.
I suspect that is the inevitable fate of any such broad-brush proclamation.

       Ethan

----------
From: Jacob Keller 

Hi Ethan, thanks for pushing me to clarify--see below.
The question is: "is there a reference in which Rmerge has been
thoroughly, clearly, and authoritatively discredited as a data
evaluation metric in the favor of Rmeas, Rpim, etc., and if so, what
is that reference?"
Let me say it clearer: when there is a conventional, standardized
method that one wants to abandon in favor of a better method, in
practice one has to make an argument for the new one and against the
old one. This is in contrast to continuing to use the conventional
method, which, even if apodictically surpassed by the newer method, de
facto needs no justification. So, in the current example, if you want
to use Rmeas or Rpim and not even report Rsym/merge, it will ruffle
feathers, even though the former is certainly superior.

Sorry for the confusion,

Jacob

----------
From: Ethan Merritt

Why assume that any of those are a valid criterion for discarding data?

I would argue that a better approach is to ask whether the data measured
in the highest resolution shell is contributing positively to the map
quality. The R_{whatever} may be an imperfect predictor of that, but is
not by itself the property of interest.

In other words, there are two separate issues in play here:

1) Is there a "best" measure of data quality in the abstract
  (i.e. it can be calculated before you solve the structure or
  calculate a map)?

2) Is there a standard statistic to choose what data is used for
  refinement?

If you just want to argue which R_{whatever} best serves to address
the first issue, carry on.

If you are worried about the second issue, IMHO none of these
quantities are appropriate.  They address entirely the wrong question.
We all know that good data does not guarantee a good model, and noisy
data may nevertheless yield a valid model. So you need a better reason
to discard data than "it's noisy".

       Ethan

> > I don't follow the logic that because a statistic is reported, one
> > must therefore argue against it.
>
> Let me say it clearer: when there is a conventional, standardized
> method that one wants to abandon in favor of a better method, in
> practice one has to make an argument for the new one and against the
> old one. This is in contrast to continuing to use the conventional
> method, which, even if apodictically surpassed by the newer method, de
> facto needs no justification. So, in the current example, if you want
> to use Rmeas or Rpim and not even report Rsym/merge, it will ruffle
> feathers, even though the former is certainly superior.


> Sorry for the confusion,
>
> Jacob
>

----------
From: Ed Pozharski

Aren't these sufficient?

Manfred Weiss & Rolf Hilgenfeld, "On the use of the merging R factor as
a quality indicator for X-ray data", J.Appl.Cryst. 30, 203-205 (1997)

Manfred Weiss, "Global Indicators of X-ray data quality" J.Appl.Cryst.
34, 130-135 (2001)

--
Oh, suddenly throwing a giraffe into a volcano to make water is crazy?
                                               Julian, King of Lemurs

----------
From: Anastassis Perrakis

Also

Nat Struct Biol. 1997 Apr;4(4):269-75.
Improved R-factors for diffraction data analysis in macromolecular crystallography.
Diederichs K, Karplus PA.

But none of these are in any way 'discrediting' Rmerge, they are just proposing more statistically sound alternatives. That is not the same ...

A.

----------
From: Jacob Keller

Yes, that's a good point--you can't really discredit Rmerge (it's just
a mathematical expression, after all, which must be translated vis a
vis redundancy), but you can show that the other R's are pleasanter
ways to represent the data.

Jacob

--