[phenixbb] R, Rfree, and wxu_scale

Pavel Afonine pafonine at lbl.gov
Sat Aug 4 19:33:09 PDT 2007

Hi Jianghai,

> Just finished a round of refinement.  Rfree went up about 1% and R 
> went down a bit.  I am wondering what happened here. 

You may want to play with wxc_scale and wxu_scale, that define relative 
contribution of X-ray term in coordinate and ADP refinement, 
correspondingly. The corresponding weights, wxc and wxu, are determined 
automatically and are good for most of the cases, but as always, there 
are some deviations (rarely significant).

Try different values for wxc_scale: just an array of values differing 
from default by say 0.25, 0.5, 1.5, 2 times and choose the one that 
gives you the best Rfree, gap Rfree-Rwork and geometry statistics. Same 
for wxu_scale. This always almost helps to resolve this kinds of issues.

> Also Rfree always went down after TLS refinement but always went up a 
> lot more after the individual ADP refinement.  Is there anything I can 
> do here? 

Here you must be careful. There are two potential traps to be aware of 
(these are not bugs but features that people have to be aware of):

1) If you do TLS first, the atoms in output file have ANISOU records. If 
you then take that  output file and refine individual ADP (without using 
TLS), by default they will be refined as anisotropic (since input file 
has ANISOU records and those atoms are considered as anisotropic). This 
is most likely not what you want and also in this case Rfree may go up 
2) If you force them to be refined isotropically, then the anisotropic 
atoms got converted to isotropic and hence both R and Rfree will go up. 
This is also you may not want to do.

So, the best, if you used TLS refinement once, keep using it in next 
refinements, otherwise you will fall into one of two traps above.

> Should I use weaker B restraints?  I tried to assign a wxu_scale 
> value.  But phenix.refine automatically adjust it to 1.81 every time.

This is intentional and always happens for TLS+ADP individual 
refinement. We found this the best by re-refining 350 models from PDB.

All these questions are explained in depth in phenix.refine 
documentation in the next upcoming PHENIX release.


More information about the phenixbb mailing list