Hi Sabine,

For instance after MR I did a bit of shaking the coordinates with pdbset (noise 0.1),

there are many ways of doing it using PHENIX:

- using the tool that is specifically designed to do this and many more of similar manipulations with your model:
phenix.pdbtools:
http://phenix-online.org/documentation/pdbtools.htm
Example:
phenix.pdbtools model.pdb sites.shake=0.1

- using phenix.refine: it can modify you model before refinement starts (so you don't have to run phenix.pdbtools):

phenix.refine model.pdb data.mtz modify_start_model.sites.shake=0.1

- by the way, 0.1 is ridiculously small. 2A or even larger is within convergence radius of phenix.refine most of the time. It depends on resolution, of course.

followed by simulated annealing in Phenix.

Cartesian or torsion?

Phenix states after SA:

Start R-work = 0.2671, R-free = 0.2992
Final R-work = 0.2312, R-free = 0.2666

When I use the output pdb of phenix directly in Refmac (with same mtz as input for Phenix)
Refmac tells me:
Initial R factor    0.2392   R free    0.2887

So I am quite puzzled about the discrepancy. Or can someone tell me if I made an error in reasoning somewhere?

There may be 10+ reasons for this, I'm sure I listed them before (about a year ago or more), so you can find it in phenixbb archives. Also, see Nat's reply. Some them:

- Fobs outliers removal in phenix.refine;
- efficient bulk-solvent and anisotropic scaling (P.V. Afonine, R.W. Grosse-Kunstleve & P.D. Adams. Acta Cryst. (2005). D61, 850-855. "A robust bulk-solvent correction and anisotropic scaling procedure") and mask parameters optimization;
- If your input data file contains Iobs (not Fobs) then different algorithms of conversion Iobs to Fobs (phenix.refine uses French&Wilson method);
... and many many many more....

In practice, I was only able obtain IDENTICAL R-factors between phenix.refine and SHELXL in a very artificial test (where bulk-solvent was turned off, I was using identical scattering factors, etc..).

Pavel.