Hi Jianghai Zhu,

However, I would try different refinement strategy, for example:
1) you do SA refinement first;
2) then you do regular refinement + TLS.

How to do SA refinement only?

To run the default protocol + SA:

% phenix.refine data.hkl model.pdb simulated_annealing=true

For more information see phenix.refine manual:
http://phenix-online.org/download/documentation/phenix/phenix/structure_refinement.html

The RMSD of bonds and angles after either SA or combined refinement are about 0.006 and 0.851 respectively. Don't you guys think the geometry restraints are too tight?

At 2.5A resolution I wouldn't worry about these values. The resolution is low enough to keep the geometry tight.

I thought the goal of the refinement for geometry is 0.02 and 2 for rmsd bonds and angle.

This is resolution dependent. At low resolution your stereochemistry rmsd should be close to ideal. At high resolution it makes sense to have the geometry more relaxed. Anyway, this should be justified by the amount of data and reasonable values of Rfree.
The automatic weight calculation in phneix.refine provides the optimal weight in most of the cases. If you like to change the default weight, you can do so as I mentioned in previous email. This is described in the Manual as well.

Another questions. After SA or refinement, the B factors vary a lot. In the same residue, B factors jump from 5 to 70. and for the next residue, B factors are from 40 to 90. The B factors don't look like well restrained. Are these numbers normal? I don't see this kind behavior of B factor in other refinement programs.

In this case you may want to increase ADP restraints contribution by decreasing X-ray term weight. To do so, decrease
wxu_scale value. By default wxu_scale=1.0. Try smaller values and see your B-factors. See Manual for more details.

Thanks for your questions !

Pavel.