[phenixbb] B factor refinement in recent versions of phenix - SUMMARY

[Pavel Afonine]

Hi,

this is just a summary..

In a nutshell the problem was following. Assuming total model structure 
factor is (for simplicity):

Fmodel ~ exp(-Boverall_scale) * exp(-Batoms)

it's clear that Boverall_scale and Batoms are correlated, and any 
combination of Boverall_scale and Batoms that maintains 
Boverall_scale+Batoms=const will not change the the total model 
structure factors (and therefore R-factors, etc).

The new bulk-solvent and overall scaling algorithm that I switched to in 
April did not care about returning isotropic component of Boverall_scale 
back to atoms:

http://www.phenix-online.org/presentations/PhenixSantaFe2012_PA.pdf

In fact (as explained above), the new algorithm does not use analytical 
form for overall scale at all.

Therefore a portion of Batoms was systematically absorbed by 
Boverall_scale. Not sure if this is a real problem, but to be consistemt 
with previous versions and expectations, recently (may be a month ago or 
so) I changed it such that isotropic exponential component of overall 
scale is added back to atoms. So current version of phenix.refine should 
not produce these "too small" B-factors anymore.

Anyone who experience this problem please try the latest Phenix from 
nightly builds:
http://www.phenix-online.org/download/nightly_builds.cgi

If this does not solve the problem please let me know. The most 
efficient way to let me know is to send me data and model files so I can 
reproduce and fix the problem.

Thanks,
Pavel

On 8/30/12 8:35 AM, "Weiergräber, Oliver H." wrote:
> Hello,
>
> I am experiencing a weird problem with B factor refinement (individual, isotropic) in phenix.refine.
>
> The dataset reveals a Wilson B factor of about 23 A². However, if phenix.refine is allowed to run for a number of macrocycles, the mean B value of the model drops continuously to about 9 A². In the end, several atoms are getting values of 0.00!
> There are no obvious abnormalities in the electron density, and R values are quite good.
> I was able to dig out an old phenix version (1.7.1) and found that this one behaves perfectly well: mean B around 26 A² (slightly above Wilson B, as is observed frequently) and minimum B around 10 A².
>
> So obviously something has happened between versions 1.7.1 and 1.8 ... The latest dev release (1143) shows the same behavior as version 1.8.
>
> I should say that this structure is somewhat peculiar (pseudomerohedral twinning plus twofold NCS), but again, the old phenix version handles that without problems.
> Moreover, the tendency for unreasonably low B factors has been observed recently for a more "conventional" structure in our lab.
>
> Any ideas what might be causing this and how to revert to the original behavior?
>
> Thanks in advance
> Oliver