On 09/16/2019 06:29 PM, Mariana Ajalla wrote:
Dear all,
We tried to use the Rfree set from a lower resolution data with a higher
resolution from the same Crystal. To do so We used aimless at ccp4i with the
option use free flag from another mtz file and extend the data.
I think it worked, but now we don't know how to be sure we have the same Rfree
set.
Does anyone have a way to prove it?
Thank you in advance,
Best,
Mariana
########################################################################
To unsubscribe from the CCP4BB list, click the following link:
https://urldefense.proofpoint.com/v2/url?u=https-3A__www.jiscmail.ac.uk_cgi-2Dbin_webadmin-3FSUBED1-3DCCP4BB-26A-3D1&d=DwIFAg&c=ogn2iPkgF7TkVSicOVBfKg&r=cFgyH4s-peZ6Pfyh0zB379rxK2XG5oHu7VblrALfYPA&m=kL1VStzNS1Omx1vfcOm7a09DSPOnVQ2zVULHRmpq3hk&s=0ECMEpulh9gbnGLzli9TlEM9bb6ZwdFGdX0v6GK9YVg&e=
Mariana,
If you solve a new dataset with a structure previously refined in an isomorphous crystal, and use
the same free set as was used for that refinement, you will generally notice that the initial Rfree
is higher than initial Rwork, as it was with the old dataset. If you start with a new free set,
Rfree and Rwork will initially be the same, and diverge with refinement. This implies that there is
some correlation between Fo-Fc calculated with the new data and Fo-Fc from the old dataset. This
would lead some to say your Rfree is "corrupted" if you choose a new set, and it will
only be valid after many cycles of refinement to "shake out" the bias. Note that Ian
Tickle qualified his results by saying it is necessary to refine to convergence for his conclusions
to hold.
How can the old and new Fo-Fc be correlated, if the errors in the new data are independent from the errors in the old dataset? Well, the errors in measurement contribute only a small fraction to the Fo-Fc differences, as evidenced by the fact that Rpim is much smaller than either Rfree or Rwork. Thinking in terms of "fitting the noise", if we define the noise to be the differences between Fo and Fc, there are two sources for that noise. One is error in measurement, and as noted these should be completely independent. The other "noise" is the difference between our best-refined structure and the actual structure- modeling solvent, fitting disorder with isotropic or even anisotropic B factors when in fact it is more complicated, single or few alternate conformations when in fact there is a range of positions, etc. If the new data actually represents the same structure, or even a new structure of the same protein in an isomorphous crystal, these errors are likely to be the same in the
n
ew dataset, and could account for the fact that the fit is better for the
working reflections which the structure has been refined to minimize against
the old dataset, when refined against the new dataset.
Again thinking of "fitting the noise", we clearly do not want to fit the noise that comes
from errors in measurement. But these are uncorrelated between the datasets, so choosing the same
free set has no effect on this. As for fitting the "noise" that is the difference between
our model and the real structure, this is a gray area. I tend to think fitting this noise (making
our structure more like the true structure) would be a good thing, and if it can be biased by using
all the reflections it would be a good thing. But the fact is it is going to make the R-free lower,
at least initially, and give you an unfair advantage over the purist who insists on using the same
free set for isomorphous structures. Until there is a general consensus on this it might be a good
idea to keep the old free set, if only because your reviewer might be one of those purists!
########################################################################
To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCP4BB&A=1
