Hi Yuri,
The protein model I am refining has 400 amino acids (3320 atoms). Some real quick calculations tell me that to properly refine it anisotropically, I would need 119,520 observations. Given my unit-cell dimension and space-group it is equivalent to about a 1.24 A complete data set.
interesting how you managed to do this... In any case this number is likely to be useless as refinement target uses both, X-ray term and restraints accounted with some weight. There are various rules of thumb that are typically specific to refinement programs you use.. Depending on data and model quality you can refine all non-solvent atoms with anisotropic ADPs starting from about 1.7-1.5A and higher. At about 1.2A and higher you can refine solvent with anisotropic ADPs as well. Again, data quality and where you currently are with refinement is important.
However, I have had a couple of cases where anisotropic B-factor refinement significantly improved R-work and R-free, while maintaining a reasonable R-gap, for lower resolution models (1.4-1.5 A, around 70,000 reflections). What is the proper way of modelling the B-factors?
Try the above rules of thumb and see what pleases the R-factors (both, work, free and gap) and makes sensible ADPs. In gray areas, like "1.7A resolution great data" or "1.4A heavily incomplete data" try multiple plausible options. That's the most robust way to find the answer. Pavel