how to treat weak data

Zbyszek Otwinowski zbyszek at mix.swmed.edu
Tue Jun 18 12:36:39 EST 1996


Jinsong Liu wrote:
> 
> I have question on how to treat weak data in the refinement. I was using X-PLOR
> in refinement, the refinement got stuck at R-free at 0.36, and R at 0.29. But
> when I looked at the map, the map looks very good, so I think the model can't
> be wrong. The statistics of R is bad, the R ( both test set and working set)
> value weak reflection is about 120%,  for the rest R is in the range 20%-40%.
> There are about 350 weak reflection which have such bad R value.
> I am afraid if I need to down-weight the weak reflection or do some thing else.
> I don't know if it is possible in X-PLOR to use a cut-off weight for weaker
> reflection. I have checked the data reduction, it seemed I can not do too much
> for it. The reason for the stuck I think is because the refinement tried to fit
> the model to those weak reflections which then bias the model.
> 

Weak reflections _always_ have large R-factor. This due to to the fact that
error is mostly in the model and error in the model results in about the
same expected error in calculated structure factor for all structure
factors (at least in one shell of resolution). From definition of R-factor
(regular or R-free) one would expect R-factor be inversly proportional to
magnitude of structure factor. This observation results in simple method
to reduce R-factor (also R-free) - omit weak reflections in the data set.
Unfortunatelly omiting weak reflections removes part of data set and results
in less accurate structure (more data never hurts).

The purpose of refinement is to get best (most accurate) structure and not
the lowest R-factor. This point is lost when weak data are omitted.
  

-- 
  Zbyszek Otwinowski                     |        zbyszek at chop.swmed.edu
  University of Texas                    |        tel : (214)-648-5098 
  Southwestern Medical Center            |        fax : (214)-648-5095 
  5323 Harry Hines Boulevard  Dallas, Texas  75235-9038
[EOB]



More information about the X-plor mailing list