I have recently been trying my hand at the ensemble refinement option in phenix.refine. I generally observe an increase in rfree after performing the ensemble refinement with a well refined (using anisotropic B) structure, but the r-free increases are fairly small. For instance, 18.53% to 18.9% (1.8 Angstrom structure) or 20.45% to 21.1% (1.9 A)...I understand what this technically means (the ensemble refined model isn't as good) but given the noise in the calculation due to the limited number of reflections for the test set (~1000), would someone consider these rfree increases significant? My apologies if this question is a well-answered one...I can't seem to find much discussion about the range of rfrees to be expected or whether a 0.3-0.5% increase is large enough to toss out the refinement.  Normally, I would take the lower rfree but because the ensemble refinement procedure gives potentially much more information, this seemed like a relevant question to ask. 

Thanks so much for your time, 

Arjun Raman

Ranganathan Lab

UT Southwestern