*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
Dirk,
The way the relative retraint and X-ray weights are set (automatically)
in SHELXL has the same affect as the scheme you describe, i.e. the model
is restrained more tightly at the beginning of the refinement and more
weakly as the agreement with the X-ray data improves. In addition there
is a command (STIR) to gradually increase the resolution as the
refinement progresses.
The idea of relaxing the restraints and then tightening them again might
indeed be a good way of getting out of local minima; see R. Stenkamp,
Acta D61 (2005) 1599-1602 for a recent example in which this enabled a
trans proline to switch to the (correct) cis peptide automatically.
George
Dirk Kostrewa wrote:
*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
Eleanor Dodson wrote:
All this is prob. true, but I repeat -
should tight default stereochemistry be a requirement at the early
stages of refinement at reso > 2.5A anyay?
Eleanor
Good question - I usually tell our PhD students to have tight geometry
(rmsd bonds between 0.012-0.015 A) at the end of the refinement run,
hoping that this reduces overfitting and model bias by at least some
extend. But your question reminds me of a different refinement protocol,
recommended in an older TNT manual: first, refine with high weights on
the X-ray term to allow larger shifts with poor stereochemistry, then do
more refinement with lower weights on the X-ray term to tighten the
geometry, again. This scheme is probably almost forgotten, but should
still be useful! A complementary approach would be to start refinement
with high weights on the X-ray term AND lower maximum resolution to
allow a larger radius of convergence, and then include higher resolution
data and tighten your geometry. This is quite labour-intensive. If I
remember correctly, BUSTER/TNT does a smooth effective resolution
limitation dependent on the quality of the fit of the model to the data
by including several sources of variances in a Luzzati-type of scaling.
As the model improves during refinement and the total variance
decreases, the effective resolution apporaches the resolution limit of
the observed data. I think, this is very elegant and should be an option
in every modern refinement package.
Best regards,
Dirk.
--
Prof. George M. Sheldrick FRS
Dept. Structural Chemistry,
University of Goettingen,
Tammannstr. 4,
D37077 Goettingen, Germany
Tel. +49-551-39-3021 or -3068
Fax. +49-551-39-2582
