On Thursday 04 February 2010 11:13:01 Ed Pozharski wrote:
Cherry-picking points to comment on, so lots of previous
discussion snipped...
> "with properly chosen restraints individual B-factor refinement is
> applicable at 3.1A resolution and, as it appears from results shown by
> Jose Antonio, may be better than two-adp-groups-per-residue refinement".
>
> > I don't see why two-B per residue wouldn't capture this distribution
> > throughout the structure (it definitely wouldn't throughout the
> > residue).
>
> Because it generates discontinuity.
Exactly. That is why I do not consider the two-B-per-residue model
to be physically meaningful, and do not recommend using it.
Think of it this way. If the number of observations is so low that
you can only afford to spend 2 refinement parameters per residue on
describing atomic displacements, what is the best way to spend those
parameters? Assigning two fixed B values per residue is, I suggest,
about the worst possible way.
A better alternative has been described by Dale Tronrud:
J. Appl. Cryst. (1996). 29, 100-104.
This approach establishes an empirical model for the increase of
B along the atoms of a sidechain, then scales this model for application
to a particular residue. You can implement this as two parameters per
residue, one base B and one scale factor for the empirical increase along
the sidechain, without introducing any discontinuities.
I am not sure if any program other than TNT offers this option.
Another alternative is to create a segmented TLS model. If you split
the chain into groups as small as 10 residues, then the 20 parameters
of each TLS description can be regarded as 2 parameters per residue.
In practice it is not necessary to make the segments that small, so
you don't need to spend even that many parameters. Note that I am
talking about a "pure TLS" model here, with no refinement of individual
Biso terms. In my recent experience using structures being refined
in the lab, these pure TLS models do as well as restrained Biso
refinement starting at roughly 2.8A resolution.
Unfortunately, if you decide to go for a segmented TLS model
there is a chicken-and-egg problem.
How to split the chain into segments for subsequent refinement?
I created the TLSMD tools to guide this choice, but TLSMD itself
needs to start with a set of refined B factors of some sort.
So for the purpose of bootstrapping the TLS model generation and
refinement, I think it is acceptable to use even crude models like
1-B-per-residue for initial refinement whose only real purpose is
to prime the TLSMD analysis.
Ethan
--
Ethan A Merritt
Biomolecular Structure Center
University of Washington, Seattle 98195-7742
