Alun R. Coker wrote:
Hi All,
I have been in the habit of transferring my initial free R assignments
to any new data sets or to isomorphous data sets such as substrate
complexes. Although theoretically this is necessary to obtain a valid
free R many of my colleagues maintain that this is completely
unnecessary in practice. Does anyone on the list have a view on this or
has anyone tested to see if it makes any difference.
Alun.
While leaving the theory to the experts, I'll suggest a test to answer
the question in an individual case. Try refining your old model against the new
data, keeping the same Free set against which the model has (not) been refined.
If the values of R and R-free are essentially the same, initially and perhaps
through rigid body refinement, then it would have been alright to pick a new
free-R set. The model is no more biased toward the working reflections (of the
new
dataset) than the free. The "noise" in the new dataset is completely independent
from the noise in the old dataset that was being "overfit" by the working
reflections.
If there is a significant gap initially, with Rfree > Rwork, then there is some
systematic error (difference) between the model and the data, which carries over
to the new dataset, which the refinement program is partially able to fit
without
improving the model as measured by the free reflections. in this case it may be
important to keep the same Free-R set, although if the model will undergo
significant rebuilding and refinement the bias may be "shaken out" before the
end.
Still, the pragmatic answer would be "just do it".
I don't think any reviewer will give you trouble for using the same free-R set.
