Hi Rongjin,
I am refining another 3.3 A resolution structure. In coot I manually adjusted the model to have >95% residues inpreferred region in ramanchandran plot, but after refinement in Phenix only <80% residues in the preferred region and a lot more in the outlier region. This was wen I used the default refinement parameters "phenix.refine mtz_file pdb_file". R and R_free are 0.22/0.27 in this case.
you may want to consider: - using secondary structure restraints; - tightening restraints (using wxc_scale=...); - using H atoms in refinement; - doing some quick geometry regularization before refinement; - fixing manually and excluding from refinement certain parts of your model. You need to experiment and find empirically which combination of the above options works in your case. I would appreciate if you share your experience with this and tell us what worked for your in the end. Pavel.