Re: [ccp4bb] Fw: Re: [ccp4bb] Twinning

[Jorge Iulek]

   Yes, I should explain better "that might not justify all your 
observations"; in more details, if such care (free set considered the twin 
law) was not taken it is probable that the gap between R and R-free is even 
larger, as Anastassis points. Might R-free become a "reasonable" value in 
this case [then there remains the problem of the huge(r) gap and rather low 
R-factor] ? John might tell us if such care as taken and then if the picture 
would change, then with new insights.
   Thanks,

J.
But, if that was the case, wouldn't you expect the opposite, Rfree to  be 
too close to R ?
here they appear to be too different !

in other words: if Rfree is not 'really free due to twin law' i would 
expect R/Rfree to be too close, not too far !

A.

On Oct 23, 2007, at 13:58, Jorge Iulek wrote:

John,

   One other point to be added, although that might not justify all  your 
observations, is to be sure that the R-free set was chosen  respecting 
the twin law. Both phenix and CNS (not automatically)  can do this. Of 
course, as pointed before, when you change your R- free set, you should 
"anneal" to avoid previous bias.

J.

----- Original Message ----- From: "Tim Gruene" <[EMAIL PROTECTED] 
AC.GWDG.DE>
To: <CCP4BB@JISCMAIL.AC.UK>
Sent: Tuesday, October 23, 2007 4:07 AM
Subject: {Spam?} Re: [ccp4bb] Twinning


There might be a few reasons for these R-values which seem to 
contradict experience:
1) The starting phases result from MR where the model is very  close and
   from a data set at higher resolution. If your 3.1A data are of
   respectable quality, the model would still correspond well to  the 
data,
   hence both R and Rfree are low (-er than the "10% rule": Expect  the
   Rwork to be around 10% of your data set resolution).

2) Your data are not complete, maybe due to the twinning

3) You messed up with your test set and entangled it to the model 
somehow,
   e.g., because you collected a second data set and did not import
   Rfree from the original one. If the model was already overrefined
   against the initial model your Rfree would come too close to the
   R-value, even though your model is wrong.

4) You might just be a very good modeller and your data, even  though 
only
   low resolution, are rather accurate.

I am sure there are more reasons. Yet, I would be very suspicious  with 
such a low Rvalue at this resolution.

John, you could try
- calculate an omit map
- remove a major part of you model, run a quick refinement and see  if 
the
  density of the removed part comes back - this should indicate  that 
you
  did not overrefine against your data
- try simulated annealing

Tim

--
Tim Gruene
Institut fuer anorganische Chemie
Tammannstr. 4
D-37077 Goettingen

GPG Key ID = A46BEE1A


On Tue, 23 Oct 2007, Anastassis Perrakis wrote:

Dear J.Kryst  -

I am basically very puzzled to see that a 3.1 A structure has R  and 
Rfree of 15.3 and 23.8 % respectively !

Is that the diffraction limit of the crystal or is it a home  dataset 
collected with rather short exposures?
Is the Wilson B very low by any chance ?

As for the gab, I am too puzzled by the absolute values to even  think 
of it. I would had never expected a 3.1 A twin dataset
to give Rfrees anywhere else than the high 20's- let alone R  below 20 
!

Tassos

PS How do you pronounce J.Kryst really?


On 23 Oct 2007, at 6:54, john kryst wrote:

Hi ccp4bb  !!

    I am refining a 3.1 A resolution structure which has partial 
twinning with twinning fraction 0.27 (space group P61).
When i refined in CNS with twin_lsq target i got 0.1529 and  0.2384 as 
R and Rfree.
There is almost 8-9% gap in R and Rfree. Map quality is good  enough 
to trace all
the atoms and ligands.

There is nothing much to build in the map too. Any suggestions  to 
reduce the gap in R and Rfree. I couldn't use shelxl because  of the 
resolution. Is it OK to stop refinement and submit the  structure.

with warm regards
John