PAfonine at lbl.gov
Tue May 5 14:31:46 PDT 2009
you can try to build a new test set of reflections using "thin shells"
algorithm. PHENIX can't create such test set, so you need to use an
outside program that can generate it. Make sure you generate the same
number of test reflections.
However, and this is a *big* however, you will have to remove the
"memory" from newly generated test set. How? Remove all waters, reset
all B-factors to average one, shake coordinates to achieve rmsd~1A and
then run 10-20 macro-cycles of refinement in phenix.refine (or until the
R-factors stop changing), that will include:
- automatic water picking;
- refinement of individual coordinates and B-factors;
- after you have done this run, run the second run of simple xyz and
b-factors refinement using "optimize_wxc=true optimize_wxu=true".
Repeat the above steps with and w/o NCS. What do you get?
I would also do the above extensive model shaking/re-refining using the
original test set, just to make sure that this will not be enough to
make the Rfree-Rwork, Rwork and Rfree change significantly without
resorting to a different test set.
PS> I also use 10% of test reflections...
On 5/5/09 2:14 PM, Maia Cherney wrote:
> Thank you Nat, Pavel,
> Actually, the NCS approaches a 3-fold crystallographic symmetry axis
> with 3 types of protein molecules plus 2 restraint groups for DNA molecules.
> The result is the same with "optimize_wxc=true optimize_wxu=true"
> options, Rfree is going up without NCS by ~1.5%. Rwork is practically
> the same.
> Should I increase the test set (now it's 5%)? or should I leave NCS=true
> for the final pdb?
> Nathaniel Echols wrote:
>> On May 4, 2009, at 6:48 PM, Maia Cherney wrote:
>>> The resolution is 2.15 A. The NCS was always on during the refinement
>>> until we got low R factors (19.2% and 21.2%). Then the NCS was turned
>>> off for the final refinement and the R factors increased, which is
>>> strange as they should be going down when you apply less restraints.
>> This isn't necessarily true - regardless of which program you use to
>> refine, R-work isn't actually the refinement target, and minimization
>> algorithms aren't foolproof either. When you remove the restraints
>> you are effectively decreasing the observation:parameter ratio, which
>> increases the risk of overfitting. If nothing else I would expect the
>> gap between Rwork and Rfree to increase, but I've also seen both
>> increase when NCS restraints were removed.
>> It may be possible that the NCS restraints are biasing your test set,
>> leading to the suspiciously low Rfree-Rwork gap, but I thought this
>> only applied when the NCS approached crystallographic symmetry, and
>> there's no such thing as fivefold crystallographic symmetry.
>> Hopefully someone else on the list can clarify this.
>> Nathaniel Echols
>> Lawrence Berkeley Lab
>> NEchols at lbl.gov
>> phenixbb mailing list
>> phenixbb at phenix-online.org
> phenixbb mailing list
> phenixbb at phenix-online.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the phenixbb