*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
On Friday 19 May 2006 02:35 pm, David W Borhani wrote:
> Does this explain why TLS refinement sometimes gives great R-factor
> statistics, but individual B's that are very low, much lower than the
> Wilson B?
The point is that the "individual B's" are not really B's at all.
They are an incremental value that is to be added to the B described
by TLS. So if you have a perfect TLS model, all of your
"individual B's" will be zero.
Except not quite, because...
> From a practical point of view (in Refmac), to have the B's drive to
> smaller values, while the U-trace goes up, is no problem...until the
> B's need to go negative (or are artificially halted at, e.g., B =
> 2.0). Problem, or not?
What we do (i.e. what the TLSMD server does) is to add an
arbitrary constant to the TLS model such that the mean deviation
from it is expected to be some positive number, rather than 0.
That gives refmac some leeway when refining these incremental
B values.
This is another point at which introducing some new convention
might be helpful. For instance, we could set this arbitrary
constant to the Wilson B. Of course then there would have to be
some mechanism for passing the Wilson B in to the TLS analysis.
Ethan
>
> There is another good reason why you should run TLSANL to produce the
> PDB file for deposition after doing a TLS refinement in REFMAC: the
> ANISOU cards it produces should in principle be unique for a given
> structure, but the TLS matrices and additive isotropic B's are not!
> You can add a constant value to these B values for all atoms in the
> same TLS group and subtract it from the diagonal elements of the T
> tensor for that group, and the fit is unchanged. In other words
> REFMAC refines one parameter too many per TLS group, and if it were
> able to do a full matrix refinement this would cause it to blow up
> with a singular matrix (Garib: please do not interpret this as a
> criticism, the REFMAC approach is a good practical solution to a
> tricky problem and we make good use of the TLS refinement in REFMAC).
>
> Possibly some future version of a refinement program will apply TLS
> restraints rather than constraints, this could be programmed in a way
> that would refine the Uij values directly and so avoid this problem.
> The deposition of restrained Uij values would then be on a par with
> the deposition of restrained atom coordinates (which we do all the
> time).
>
> George
--
Ethan A Merritt
Biomolecular Structure Center
University of Washington, Seattle WA
