[phenixbb] high average b-factor vs. Wilson B - EXPLANATION

Pavel Afonine pafonine at lbl.gov
Fri Aug 6 10:43:32 PDT 2010


  Hi Sue,

> It seems to me that it means that the refinement strategy isn't stable

in summary, in that email I wrote that there are two ways of how you can 
handle the total B-factor and one of its components (Bcryst). You can 
subtract the trace of Bcryst and add it to individual ADPs (what CNS 
does, for example) or you can keep Bcryst as it is. Both ways do not 
change the Fmodel and therefore the R-factors. So I can't see how you 
conclude from this about refinement stability. These are purely 
formatting/convention things. Regarding the physics of this phenomenon 
(why truly total ADP (and not "the total ADP minus Ucryst") may be 
higher than Wilson B), Paul Adams just commented on this which I 
copy-paste here and which I totally agree with:

"""I think it might be worth looking at the distribution of ADPs in the 
structure. Is there one part of the structure with a average around 27 
and another part significantly higher. The problem with the Wilson 
analysis is that, I think, it will tend to give you the average B-factor 
for the best diffracting part of structure. However, when you solve the 
structure and calculate the average B-factor it will be across the whole 
structure including the more disordered parts. Hence the average B from 
the structure could be higher than the observed Wilson B."""

Regarding the refinement stability... I think I mentioned this a few 
times before. Just repeating: If you run 100 identical refinements where 
the only difference is the random seed you will get an ensemble of 
slightly different results, and the difference is more significant as 
lower the resolution you have. At low resolution you may end up with 
1-2% difference in R-factors... See slides #9-10 here:
http://www.phenix-online.org/presentations/latest/pavel_validation.pdf
that illustrate the grounds for this.
Why random seed makes it different? This is because the X-ray/Restraints 
weight calculation is done after some random shaking of the coordinates, 
and therefore the weight can come out slightly different, which is 
enough for refinement to take a different path to a different local 
minimum. Nothing magic here.

Similar logic applies to the ways of handling total ADP and Bcryst. By 
doing one way or another, numerically it might provide enough of 
difference to result in sightly different R-factors.

> the contributions from the overall Baniso and the individual B factors are not being (and prossibly can not be) separated properly

Decomposing total B-factor

Utotal = Ucrystal + Ugroup + Ulocal,

where Ugroup = Utls + Ulib (see recent PHENIX Newsletter for definitions)

into individual components properly is much complex problem than it may 
appear. An attempt to solve it partially for a specific case is sketched 
here (especially page 9):

http://www.ccp4.ac.uk/newsletters/newsletter45/articles/Tmax-CCP4_Afonine-Urzhumtsev_2006_arial.pdf

Do you know a proper (or at least better) solution for a more general case?

All the best!
Pavel.




More information about the phenixbb mailing list