Next: Results Up: Knowledge-Based B-Factor Restraints for Previous: The Proposed Method
Next: Results
Up: Knowledge-Based
B-Factor Restraints for Previous: The
Proposed Method
Before these new restraints can be applied to a model the values of
and 
must be determined from accurately-known structures in which we have confidence.
The choice of the basis set is complicated because on one hand one would like these models to be based on very high resolution diffraction data sets, but on the other hand the motion of the atoms in the models should be similar to those in ``typical'' macromolecules. In particular, one would not expect the motion seen in crystal of small molecules to be representative of the motions of macromolecules.
The values of
and
must therefore be derived from protein molecules. The resolution of the
diffraction data must be fairly high, e.g. at least 1.7Å. The B-factors
of the models should not have been restrained by any factors, such as those
that one is attempting to derive. This restriction prevents any bias caused
by either the traditional or new restraints. Because the exclusion of low
resolution diffraction terms cause systematic errors in the B-factors,
the models used to derive the standard values must have been refined against
all of the data.
Since most models produced by refinement with PROLSQ or XPLOR are routinely subjected to B-factor restraints, they cannot be used to determine standard values for the new restraining function. This excludes a very large proportion of the structures in the Protein Data Bank.
Therefore it was necessary to resort to a set of four structures, with a total of about 900 amino acid residues. The structures are listed in Table 1. The library of standard values was generated by calculating the mean B-factor change for each type of bond in all the models, as well its standard deviation.
Parameters for nucleic acids have not been defined due to the lack of a basis set.
A benefit of this survey is that it provides confidence limits for the
B-factors derived by the refinement procedure. If the B-factors for these
models were unreliable all of the
parameters would be quite large. In fact, the mean
for bonds between atoms that are well ordered (e. g. main chain
atoms and the side chain atoms of hydrophobic side chains) averages about
5Å 
. This implies that the 95% confidence interval for the B-factors of these
models is approximately 
Å
( 
because the integral of the Normal distribution from -1.96 to 1.96 is 0.95).
