As always, Dale hits the nail on the head.
Your R factor is a function of your scaling algorithm as well as your
model, and as you point out the apparent R factor increases when you
include the low resolution data. This doesnt mean your model is worse -
it is still the same model!
All refinement programs now have a fudge correction to try to account
for the missing solvent model. This usually assumes that the low
resolution data distribution is OK, ie the data is a) measured and b)
measured correctly..
Unfortunately this is often not so - all strong low resolution
reflections may be overloads and therefore missing altogether giving a
very distorted distribution.
(Look at any of the plot from CCP4 data processing or refinement which
show <F> v resolution. )
And if the scaling algorithm gets itself in a twist it can force up the
R factors across all the resolution ranges - again you can check a
REFMAC plot of <Fobs> and <Fcalc> v resolution. They should overlap but
if the scaling is poor they may not..
Eleanor
Dale Tronrud wrote:
Dear Bert,
I think is it important here to remember that R and R free are tools
to help judge the quality of the model, not the data. You cannot reject
data because it disagrees with your model. Lowering your free R by
changing your resolution limit is no different than finding all the
individual reflections that disagree most with your model and rejecting
them. Neither procedure increasing the quality of your model, it
simply lowers the R value.
If you suspect you have a problem with overloads, you have to go
to your data reduction log files and see if you have this problem. If
you suspect your bulk solvent model is poor you need to find out what
model you are using and decide if it is correct for your crystal. It
is certainly a good idea, as Pavel suggests in another response, to
try other programs that have different approaches.
Membrane protein crystals have a particular problem with bulk solvent
because there is the lipid part of the bulk solvent and the water part
and most bulk solvent models assume a single kind of bulk solvent.
Hopefully someone else will reply with some references to such
two-component bulk solvent models.
If you can't figure out how to lower your R's by improving your
model, I'm afraid you just have to live with it. Your R values will
be higher because they reflect the inability of your model to fit
your data.
Dale Tronrud
Van Den Berg, Bert wrote:
Hi all,
during refinement of our (membrane protein) structures, basically in
all cases the R/Rfree values depend a lot on the low resolution
cutoff. Putting the cutoff at lower res (20-50 A) results in
substantially higher R/Rfree values (sometimes few percent). For this
reason we mostly refine the data from the high-res limit down to 10A
or so. I have noticed that this occurs fairly often in the
literature, but I don't know if this is a membrane protein related
issue or not.
Could it be that the bulk solvent model used in CNS (we refine
exclusively with CNS) does not model the situation with membrane
proteins, due to the presence of detergents? Or is it related to data
collection issues (low-res spots overloaded etc)? Anything else? What
could be done to overcome the problem, and to use all the data in
refinement?
Thanks, Bert
Bert van den Berg
University of Massachusetts Medical School
Program in Molecular Medicine
Biotech II, 373 Plantation Street, Suite 115
