[phenixbb] refinement question

[Maia Cherney]

Hi Pavel,
Thank you for your letter. You mentioned "unless it is highly systematic 
and consistent observation .."
Yes, it is highly systematic and consistent for these data. I continue 
to refine this structure with both strategies (with  and without ncs). 
Weights optimization keeps the same difference between Rfee factors 
(improves both). I liked the suggestion by Engin to keep the ncs, but 
remove some offending residues from the ncs restraints. (There are such 
residues that have slightly different conformations and they get moved 
out of density by the ncs restraints).

Maia



Pavel Afonine wrote:
> Hi Maia,
>
> I would add a couple of comments too:
>
> 1)
>   
>> I am wondering why the ncs refinement gives me a better Rfree (21.0%) 
>>     
>
> Because by using NCS you added some "observations", therefore you 
> improved the data-to-parameters ratio, which in turn reduced degree of 
> overfitting.
> The fact that the R-factor dropped (and not increased) probably suggests 
> that you selected NCS groups correctly.
>
> 2)
>   
>> the ncs refinement gives me a better Rfree (21.0%) 
>> than without ncs (21.7%). 
>>     
>
> Here is another stream of thought...
> The target for restrained coordinate (and similarly for B-factor) 
> refinement looks like this (in phenix.refine):
>
> T_total = wxc_scale * wxc * T_xray + wc * T_geometry
>
> where the relative target weight wxc is determined as  wxc ~ ratio of 
> gradient's norms:
>
> Brünger, A.T., Karplus, M. & Petsko, G.A. (1989). Acta Cryst. A45, 
> 50-61. "Crystallographic refinement by simulated annealing: application 
> to crambin"
> Brünger, A.T. (1992). Nature (London), 355, 472-474. "The free R value: 
> a novel statistical quantity for assessing the accuracy of crystal 
> structures"
> Adams, P.D., Pannu, N.S., Read, R.J. & Brünger, A.T. (1997). Proc. Natl. 
> Acad. Sci. 94, 5018-5023. "Cross-validated maximum likelihood enhances 
> crystallographic simulated annealing refinement"
>
> Before wxc scale is computed, the structure is subject of a short 
> molecular dynamics run - this is where the random  component comes into 
> play.
>
> Now, having said this, we know that if you run, for example, 100 
> identical phenix.refine runs, where the only difference between each run 
> is the random seed, you will get 100 slightly different refinement 
> results. The spread in R-factors depends on resolution, and if I 
> remember correctly, for a structure at ~2A resolution I was getting 
> delta_R~ from 0.1 to 2%. It can be higher at lower resolution, and 
> smaller at higher resolution.
>
> The NCS term goes into T_geometry, which in turn means that it changes 
> (somehow) the weight. This may explain the difference in R-factors and, 
> I would say, the one less then 1% I would consider insignificant (unless 
> it is highly systematic and consistent observation, and unless it is not 
> made weight independent).
>
> To make it less arbitrary I would suggest to run two refinement jobs 
> using "optimize_wxc=true optimize_wxu=true", one with NCS and the other 
> one without using NCS. I'm sure I suggested this a month or two ago.
>
> Pavel.
>
>
> _______________________________________________
> phenixbb mailing list
> phenixbb at phenix-online.org
> http://www.phenix-online.org/mailman/listinfo/phenixbb
>
>