On Fri, Jul 11, 2014 at 8:15 AM, Chen Zhao <c.zhao@yale.edu> wrote:
1) R-free flag in refinement
a) For twin refinement, I know the reflections related by the twin law should have the same R-free flag. If I process my data in P1, and generate the R-free flag by phenix.refine using the key word xray_data.r_free_flags.use_lattice_symmetry=True, can phenix generate the R-free flags considering the 2-fold point group symmetry and deal with the twinning properly?

In theory, yes (and FYI, that parameter is already True by default).  You can check for this by running the program phenix.explore_metric_symmetry, providing your symmetry as input parameters, and it will show you what it thinks the lattice symmetry is.  If it says C2, that's what phenix.refine will do too.

b) I read that the R-factor could go down to around 40% in twin refinement in the presence of twinning even if the model is completely wrong. For my understanding, this could be due to the fact that the reflections are correlated by the twin law. If the R-free flags are chosen in the highest lattice symmetry, will the R-free still suffer from this problem?

Using the lattice symmetry avoids the biasing of twin-related reflections.  The phenomenon of lower R-factors for poor models in twin refinement is something different - since I am not an expert at statistics I am not going to try to explain it myself, but here is a paper by Garib Murshudov that gives the mathematical derivation (which is independent of implementation of the refinement program):

http://www.ysbl.york.ac.uk/refmac/papers/Rfactor.pdf

2) Twin fraction
In phenix.xtriage, the predicted twin fraction is about 30%. However, in phenix.refine, the twin_fraction is always between 46% to 50%. Why could this happen?

The prediction is just that - a guess, based on the experimental data alone.  (And there are several different predictions, which won't all give the same answer.)  The twin fraction given by phenix.refine actually takes the model into account, so it should be more accurate.  Although I think you should be incredibly careful when dealing with data that are (almost) perfectly twinned; the combination of low resolution and detwinning means that the 2mFo-DFc map will be nearly useless.  I forget whether there are other things to check for when perfect twinning is indicated by refinement - occasionally this is diagnostic for other problems, errors in data processing, etc., but I've never figured out the exact rules.

3) Data reduction
I am not very clear about this point, since I just started dealing twinning. But what I am thinking is whether it is a good idea to merge the Friedel pair in the presence of twinning. Since the two twin domains have different hkl index if considering one as the reference, is the Friedel law still valid? This might be a very simple mathematical problem.

The twin law should apply to Friedel mates as well, so F+ and F- will be combined with F_twin+ and F_twin-.  Unless you have enormous anomalous differences, merging them should not be a problem.  You probably won't be able to get much useful anomalous information out them anyway.

-Nat