Dear Randy,
thanks for your comment - a good point with the likelihood target
estimated from E-values! So, in principle, there shouldn't be any
difference in maximum-likelihood refinement using sharpened data or not.
However, for curiosity, in one case at 4.3 A resolution and a sharpening
B-factor of ~100 A**2, I compared ML refinement against the sharpened
data with ML refinement against the original data and subsequent map
sharpening: the R-factors were almost identical, and so were the
electron density maps in most places. But in a some places, the maps
were slightly different, with slightly less (!) model-bias and slightly
clearer densities for the refinement against sharpened data. But those
judgements were very subjective, and since a "true" structure at really
high resolution is not available, I never could quantify this. Either,
this was only anecdotal evidence, or there is still room for improvement
in existing ML refinement programs.
Best regards,
Dirk.
Am 28.02.11 11:40, schrieb Randy Read:
Hi,
I'm on Garib's side here. The way the maximum likelihood targets work, the
variances are defined relative to the average intensity in a resolution shell,
so if you change the falloff the variances will change in the same way. In
fact, one way to implement maximum likelihood refinement is in terms of
E-values, from which the falloff has been removed. If your B-factors didn't
run into hard limits (which, as Garib points out, they will when you make the
data non-physical) you would end up with the same model if you refined against
sharpened data, except the B-factors would be lower. The other thing that will
change if you sharpen the data is that the R-factors will be higher, because
the poorly-fit higher-resolution terms will contribute more to the sums. And
that's probably not what you want when you might already have a hard time
getting a low resolution structure past the freeR police!
This is a case where intuition can lead you astray. Intuition might suggest
that, if you sharpen the data, the refinement program should pay more attention
to fitting the high resolution detail, but the likelihood target doesn't look
at the data the same way you do when you look at a map. The fact that you can
define the target in terms of E-values means that, if your model and data are
both good, the likelihood target can be thought of as sharpening the data
anyway.
Best wishes,
Randy Read
On 28 Feb 2011, at 09:02, Dirk Kostrewa wrote:
Dear CCP4ers,
I really would sharpen the structure factors, not only the electron density
maps. The simple reason is: if sharpening emphasizes enough information at
higher resolution to help interpreting the electron density maps, refinement
will also benefit from this information.
Of course, the mean B-factor of the refined structure will be lower by the
sharpening B-factor, but since B-factor sharpening is usually done with lower
resolution data, the Wilson B-factor is usually very high, and thus far, I
didn't run into problems with B-factors crashing at the lower limit.
The sharpening B-factor can easily reach values in the -100s A**2, not only -10
to -50 A**2. Axel Brunger has published several papers about how to estimate a
good sharpening B-factor (a recent one with references is Brunger et al. Acta
Cryst D65, 128-133). He usually describes map sharpening, but B-factor
sharpening of structure factors seems to be done routinely for virus structures
in Steve Harrisons lab.
One word of caution: the B-factor sharpening should be correctly described not
only in the publication but also in the PDB deposition (if refinement was done
against sharpened structure factors, the refinement statistics can only be
reproduced using these structure factors). The original structure factors can
be easily reproduced by applying back the negative sharpening B-factor.
Best regards,
Dirk.
Am 26.02.11 01:09, schrieb Garib N Murshudov:
I would not sharpen structure factors before refinement. It may cause problems
with B value refinement (a lot of B values may stuck around 2 or minimum B).
One must remember that not all atoms in crystal have same Bvalue. There is a
distribution of Bvalues.
However maps can be sharpened after refinement. It can be done directly in coot
(I hope this version of coot is now widely available). Or if you are using
refmac for refinement you can use:
mapc sharpen # regularised map sharpening. Bvalues and regularisation
parameters are calculated automatically
or
mapc sharpen<Bvalue> # regularised map sharpening with specified Bvalue
or
mapc sharpen<Bvalue>
mapc sharpen alpha<value=0.1> # regularisation paramater. alpha=0 is simple
sharpening.
I am sure other programs have similar options. (I know CNS has and it has been
used successfully by many people)
regards
Garib
P.S. These options available from refmac v5.6 available from;
www.ysbl.york.ac.uk/refmac/data/refmac_experimental/refmac5.6_linux.tar.gz
On 25 Feb 2011, at 23:57, Dima Klenchin wrote:
At 05:39 PM 2/25/2011, Pete Meyer wrote:
Or could anyone suggest a program that would be of help?
CAD scaling with a scale factor of 1.0 and negative B-factor (isotropic or
anisotropic) should do the trick. I haven't had much luck with density
sharpening (at least at ~4-5 Angstroms), but others have apparently had some
success with it.
Alternatively, CCP4i task Run FFT does the job:
1. Take MTZ from Refmac output
2. Run FFT to create simple map with SigmaA-weighted phases (i.e., PHWT label).
3. In Infrequently used options, "Apply B-factor scaling to F1", specify
negative B-factor scaling value, usually within -10 to -50.
- Dima
--
*******************************************************
Dirk Kostrewa
Gene Center Munich, A5.07
Department of Biochemistry
Ludwig-Maximilians-Universität München
Feodor-Lynen-Str. 25
D-81377 Munich
Germany
Phone: +49-89-2180-76845
Fax: +49-89-2180-76999
E-mail: kostr...@genzentrum.lmu.de
WWW: www.genzentrum.lmu.de
*******************************************************
------
Randy J. Read
Department of Haematology, University of Cambridge
Cambridge Institute for Medical Research Tel: + 44 1223 336500
Wellcome Trust/MRC Building Fax: + 44 1223 336827
Hills Road E-mail: rj...@cam.ac.uk
Cambridge CB2 0XY, U.K. www-structmed.cimr.cam.ac.uk
--
*******************************************************
Dirk Kostrewa
Gene Center Munich, A5.07
Department of Biochemistry
Ludwig-Maximilians-Universität München
Feodor-Lynen-Str. 25
D-81377 Munich
Germany
Phone: +49-89-2180-76845
Fax: +49-89-2180-76999
E-mail: kostr...@genzentrum.lmu.de
WWW: www.genzentrum.lmu.de
*******************************************************
