Hi,
here is a question for the crystallography jocks out there.
Stimulated by Sue Roberts' email here is a thought experiment that I
have puzzled about
in the past. Basically I think that if one has a wrong structure and
then assumes the
crystal is twinned one will get a lower R-factor.
Not having actually refined a twinned structure in my life I may be
off the mark,
so please feel free to tell me that I am.
Lets assume you have a data set and a structure and the two have
nothing to do with one
another – other than that they belong to the same space group and
have the same unit cell dimensions.
The R-factor in such a situation will then be the random R-factor,
which is somewhere around 0.6 (but
you can probably get it down to somewhere around 45% by refining).
Now if you declare your data to be twinned and calculate the Fcalc's
from the twinned structure,
these Fcalc_twin will have a narrower distribution of amplitudes
(from what I understand, this
is how one spots twinned data in the first place). I would then
expect the average discrepancy (and with
it the R and R-free) between the Fobs and the Fcalc_twin to be
smaller. This lowering of the R-factor may then
give the false impression that the structure was indeed twinned even
when it is not.
Another way of thinking about it
is that the twinning essentially blurs the structure, such that it
does not disagree too badly with the true
structure. The net effect would be similar to letting the R-factor go
through the roof.
Does this make sense? And if it does, what is the drop in R and R-
free one would expect purely
based on allowing a random structure to be refined as twinned.
Cheers,
Ulrich
On Feb 21, 2007, at 10:22 AM, Sue Roberts wrote:
> Hello
>
> A partially philosophical, partially pragmatic question.
>
> I've noticed a trend, both on ccp4bb and locally, to jump to
> twinning as an explanation for data sets which do not refine well -
> that is data sets with R and Rfree stuck above whatever the
> person's pre-conceived idea of an acceptable R and Rfree are.
> This usually leads to a mad chase through all possible space
> groups, twinning refinements, etc. and, in my experience, often
> results in a lot of time being spent for no significant improvements.
>
> Just out of curiosity, does anyone have a feel for what fraction
> of stuck data sets are actually twinned? (I presume this will vary
> somewhat with the type of problem being worked on).
>
> And a sorta-hypothetical question, given nice-looking crystals;
> images with no visible split spots, extra reflections, or streaks;
> good predictions; nice integration profiles; good scaling with
> reasonable systematic absences; a normal solvent content; and a
> plausible structure solution, and R/Rf somewhat highish (lets say .
> 25/.3 for 1.8 A data), how often would you expect the Stuck R/Rf
> to be caused by twinning (or would you not consider this a failed
> refinement). (My bias is that such data sets are almost never
> twinned and one should look elsewhere for the problem, but perhaps
> others know better.)
>
> Sue
> Sue Roberts
> Biochemistry & Biopphysics
> University of Arizona
>
> [log in to unmask]
|