The discussion is converging, but let me clarify a few points.  I suspect
that a new generation of crystallographers has come to the fore since the
last material discussions of real-space refinement in the mid 1990's.  So
with apologies to the grey-haired (like me) set who may find the following
condescending...

Fourier transform theory and the experience of Jack/Levitt led to the
argument that real- and reciprocal- space refinements were formally
equivalent.  It can be true, but is only an approximation under the
conditions of practical refinements. *Close to convergence*, the
approximation is better - reciprocal space refinement is better without the
implicit real-space restraint to the current phases.  Improvements such as
sigmaA, maximum likelihood etc. have extended the domain where reciprocal
space is better, but the trade-offs depend on data, model quality.

The term real-space refinement is commonly used for two different
proceedures.  In "true" real space refinement, model density is matched to
experimental density directly.  In what is also called real-space
refinement, but should more properly be called pseudo-real-space refinement,
the refinement is actually performed in reciprocal space, either by matching
model and experimental structure factor vectors or by adding explicit phase
restraints.  As hopefully evident below, pseudo-real-space refinement has
only 1/2 the advantages of true real-space refinement.

The "real" value of real-space refinement is in reducing over-fitting &
avoiding false minima, challenges that loom bigger earlier in refinement.
Real-space refinement can do this in two ways:
1.  Increased restraints from experimental phase information.  Note that
this is no longer an advantage at stages of refinement when experimental
phase information has been (completely) discarded.  Both true and pseudo
real-space methods can use phase information.  Pseudo real-space methods can
have an advantage in that the phase information can be weighted differently
from the amplitude information.
2.  Localizing refinement.  As Brunger's group showed, much over-fitting
results from mutually compensating errors / omissions from different parts
of a structure.  True real-space refinement avoids this, because the
optimization is against local density values, not structure factors
dependent on the entire structure.  Pseudo real-space refinement is just as
susceptible to over-fitting errors as reciprocal-space refinement, because
both are "global" refinements.  

One could say that we all alternate real-space and reciprocal-space
refinements with our alternated rebuilds / refinements.  Early in refinement
or at low resolution, automated real-space improvement is a worthwhile
additional step.  Over-fitting on reciprocal-space refinement is much less
with a starting R=0.35 model than the R>0.45 that I'm capable of by manual
building.

There are several programs that can be used for this real-space refinement.
They differ in model parameterization: all atom and/or rigid fragment;
cartesian vs. torsion angle; optimization method; objective function: true
vs. pseudo; Gaussian atoms vs. resolution-dependent form factor calculation.
In many cases, these distinctions may not be critical, and convenience of
use may drive the decision of which to use.   At lower resolution and with
crude initial models it may be important to use a "full-feature"
implementation, such as RSRef which can combine real-space refinement with
the reduced parameter Rice/Brunger torsion angle SA/CG optimizer in
CNS/X-plor.

Bottom line messages:
1.  Real-space refinement is usually a worthwhile prelude to
reciprocal-space refinement.  (Importance tends to rise with the difficulty
of structure determination...)
2.  "Real-space refinement" is a term used loosely.  Dig deeper to
understand the algorithms in use in your favorite programs, and make an
appropriate selection based on your task at hand.

Michael.

Michael S. Chapman, R.T. Jones Professor of Structural Biology
Dept. Biochemistry & Molecular Biology; School of Medicine, Mail Code L224
Oregon Health & Science University
3181 Sam Jackson Park Road; Portland, OR 97239-3098
[EMAIL PROTECTED] / (503) 494-1025; http://xtal.ohsu.edu/

Reply via email to