[phenixbb] R/Rfree from phenix.autobuild

[Pavel Afonine]

Hi Graeme,

the difference between Rwork and Rfree depends on resolution. Here is 
some R-factor statistics for all PDB structures refined at resolution 
range between 1.55A and 1.65A:

Distribution of R-work at resolution 1.55-1.65
 Resolution range  Number of structures
  9.90 -  11.86:     6
 11.86 -  13.82:    28
 13.82 -  15.78:   185
 15.78 -  17.74:   415
 17.74 -  19.70:   562
 19.70 -  21.66:   358
 21.66 -  23.62:   144
 23.62 -  25.58:    27
 25.58 -  27.54:     8
 27.54 -  29.50:     2

Distribution of R-free 1.55-1.65
 Resolution range  Number of structures
 14.00 -  15.93:    11
 15.93 -  17.86:    73
 17.86 -  19.79:   293
 19.79 -  21.72:   490
 21.72 -  23.65:   458
 23.65 -  25.58:   244
 25.58 -  27.51:   110
 27.51 -  29.44:    39
 29.44 -  31.37:    11
 31.37 -  33.30:     6

Distribution of R-work-Rfree 1.55-1.65
 Resolution range  Number of structures
  0.10 -   1.23:    69
  1.23 -   2.36:   426
  2.36 -   3.49:   562
  3.49 -   4.62:   401
  4.62 -   5.75:   163
  5.75 -   6.88:    67
  6.88 -   8.01:    31
  8.01 -   9.14:    11
  9.14 -  10.27:     3
 10.27 -  11.40:     2

1% difference in your case is suspiciously small indeed. Although 
Autobuld combines model building with refinement using phenix.refine, 
its main purpose to build the structure and not produce the best refined 
possible one. After Autobuld you may want to do some more refinement 
customized to your particular case (resolution, etc...). At 1.6A you may 
want to try:

- refining anisotropic ADP;
- if refinement of anisotropic ADP is prohibitive (Rfree and Rfree-Rwork 
go too high), try TLS instead;
- simply do more refinement macro-cycles to achieve full convergence (by 
default phenix.refine does 3 macro-cycles), say try 5, for example;
- optimize weights using automatic options (use "optimize_wxc=true 
optimize_wxu=true" - this may take a while to finish), or try a few 
different values manually (wxc_scale=..., wxu_scale=...).

Please let us know if you have any other questions!

Pavel.


On 11/27/2008 8:54 AM, Graeme Winter wrote:
> Hi Folks,
>
> Quick question about R/Rfree from phenix - I ran the autobuild and it
> kindly built pretty much all of the visible residues, and it gave me
> the following output:
>
> SOLUTION  CYCLE     R        RFREE     BUILT   PLACED
>  1         1      0.22        0.23      169       169
>  2         2      0.21        0.23      170       170
>  3         4      0.22        0.23      170       170
>
> Now, I'm not one to complain about good results, but I did feel that
> the Rfree looked a little close to the R factor - usually I would
> expect a difference of maybe 4-5%. Is there a reason for this that I
> am missing? The model seemed to fit very nicely, closely corresponded
> to the PDB entry and the waters were pretty good. There is no NCS (one
> copy in the ASU) and the phases are from SeMet mad data from Autosol.
>
> Resolution is ~1.6A, if that is any help.
>
> Just noticed this and thought it was interesting. I am used to
> Arp/wArp for automated tracing, which typically gives a bigger
> difference...
>
> Many thanks,
>
> Graeme
> _______________________________________________
> phenixbb mailing list
> phenixbb at phenix-online.org
> http://www.phenix-online.org/mailman/listinfo/phenixbb
>   
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://phenix-online.org/pipermail/phenixbb/attachments/20081127/9fbc3680/attachment-0003.htm>