Hi Sam,

to help with questions like yours, recently we have developed and implemented in PHENIX a new tool called POLYGON. See
Acta Cryst. (2009). D65, 297-300
for details.

I'm refining a 2.88 Angstrom structure and am getting a difference of greater than 6% between the R-work and R-free.
(...)
The best result (I think) is what I have from minimization alone:
Final R-work = 0.2411, R-free = 0.3176

Here is the distribution of Rwork and Rfree for "all" structures in PDB refined at resolutions between 2.78 and 2.98A, that is similar to yours, and "<<<" indicates where your "best result" is standing with respect to this distribution:

r_work_pdb data_points=1411 min/max/mean= 0.1860 0.2870 0.2252
  0.1860   0.1961 121
  0.1961   0.2062 190
  0.2062   0.2163 207
  0.2163   0.2264 233
  0.2264   0.2365 249
  0.2365   0.2466 175 <<< your structure
  0.2466   0.2567 109
  0.2567   0.2668 67
  0.2668   0.2769 42
  0.2769   0.2870 18

r_free_pdb data_points=1411 min/max/mean= 0.1940 0.3400 0.2744
  0.1940   0.2086 4
  0.2086   0.2232 28
  0.2232   0.2378 73
  0.2378   0.2524 163
  0.2524   0.2670 234
  0.2670   0.2816 363
  0.2816   0.2962 298
  0.2962   0.3108 149
  0.3108   0.3254 71 
<<< your structure
  0.3254   0.3400 28

It seems like your Rwork is reasonably good, and although there are more than 100 models with similar or worse Rfree, I believe you can push it below 30%.

Things to try:

- TLS. Use TLSMD to define TLS groups.
- Use "optimize_wxc=true optimize_wxu=true" keywords. This may take longer to run.
- Although 2.88A is pretty low for automatic water picking, I would still give it a try.
- Torsion angle dynamics.
- Try the four above with and without NCS. Let phenix.refine automatically defined NCS groups, but manually check is the automatic choice makes sense.

Once done with trying and obtained the best result, use the following tools for validations:

phenix.real_space_correlation
phenix.polygon
phenix.model_vs_data

or phenix.validation
where the three above are included along with the cool things from Molprobity team. You need a recent version of PHENIX for this with the new GUI:
http://www.phenix-online.org/download/nightly_builds.cgi

Please let me know if you have any questions!

The results from simulated annealing are the following:

Final R-work = 0.2403, R-free = 0.3333


Yes, both Rfree and Rfree-Rwork seem pretty large.

I have also done the refinement with ncs but ncs is probably not appropriate for my structure.  The final R-free ends up being higher (though the difference between R-free and R-work ends up being 6-7%)

This may indicate that the NCS groups were not optimal or TLS is needed to NCS-restrain only residual B-factors, or you need to optimize the weights (see above).

Once again, please let me know if you have any questions.

Pavel.