*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
Dear Clemens,
I can only second your statement. For (almost) atomic resolution
structures, there are enough observed data to allow refinement without
any ideal geometry parameters or with only little weights on them,
resulting in RMSDs comparable to well refined small organic molecules
(i.e. RMSD bonds ~ 0.02 A). For lower resolution structures, there are
not enough observed data and one has to use ideal geometry parameters as
additional observed data. The refinement against Fobs _and_ ideal
geometry results in RMSDs lower than the ones in the ideal parameter set
(see Cruickshank, Acta D55, 583-601, 1999).
My personal advice for non-atomic resolution structures is to set the
relative weights such that the refined RMSD for bonds is lower than 0.02
A, and I recommend values between 0.010-0.015 A, with lower values for
lower resolution and vice versa.
Best regards,
Dirk.
Clemens Vonrhein wrote:
*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
Dear all,
maybe I'm missing the point completely, but aren't we comparing apples
and pears here? What we have in this type of argument is
1. "RMSD on bond lengths"
I guess this is what we have in the PDB headers as
RMS DEVIATIONS FROM IDEAL VALUES
and what is defined in mmCIF as
_refine_ls_restr.dev_ideal
For the given parameter type, the root-mean-square deviation
between the ideal values used as restraints in the least-squares
refinement and the values obtained by refinement. For instance,
bond distances may deviate by 0.018 A (r.m.s.) from ideal values
in the current model.
See:
http://mmcif.pdb.org/dictionaries/mmcif_std.dic/Items/_refine_ls_restr.dev_ideal.html
2. the Engh & Huber set of parameters
These are given as a distribution (mean and sigma) for different
types of bonds (Table 2 in paper):
"Averages, standard deviations and standard errors were
calculated, with and without elimination of values outside four
standard deviations from the mean."
See:
http://journals.iucr.org/a/issues/1991/04/00/li0061/li0061.pdf
These two quantities are something completely different - at least as
far as I understand it. The rms deviation from the mean of a
distribution is only identical to the standard deviation of that
*same* distribution.
I don't understand why the RMSD on bond lengths should/can be compared
to the standard deviation of the Engh&Huber distributions to see if a
structure is well restrained?
If we would calculate the distribution of bond lengths within a
refined structure, we would also get a mean and a sigma. And this
could be compared to the Engh&Huber parameters, since it measures the
same thing.
However, there is some sense in arguing about well restrained
structures while looking at the RMSD on bond lengths:
- high resolution => more observations than parameters => the data
tells us what the bond-lengths should be => we can deviate from
the Engh&Huber mean values => larger rmsd(bond)
- low resolution => few observations => the data doesn't tell us
what the bond-lengths should be => we can't really deviate from
the Engh&Huber mean values (prior) => smaller rmsd(bond)
But what should be those values? I guess at 1A resolution I'd expect a
rmsd(bond) of 0.02 or larger. And at 3.5A I'd expect a rmsd(bond) of
0.005 or smaller. But these are personal choices (same as I/sig(I)
or Rmerge cut-offs in data-reduction).
But maybe I'm completely wrong here ... correct me please?
Cheers
Clemens
On Wed, Aug 23, 2006 at 09:50:04AM +0200, Fred. Vellieux wrote:
*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
On Tue, 22 Aug 2006, Bernhard Rupp wrote:
*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
at the resolution you're working at the RMSD on bond
lengths should be ca. 0.012, not 0.02 A.
Interesting statement. I agree that for
very well determined small molecule structures, the
overall bond length variation is about 0.014 A or so.
That fits quite well with what you say. It just means
that the structure should be restrained to reflect
reality.
But: What does coordinate rmsd exactly have to do with
resolution? If the 3.5 A structure has 0.012 A
rmsd, does that mean that a 1.2 A structure
should have 0.005? I have a feeling that would
be a serious case of over-restraining.
Should not all structures reflect the same
'real world' rmsd (plus minus some minor individuality)
if the restraint weights are properly selected?
Maybe the small molecule/shelxl fellows may chime in here:
What is the expected rmsd for a high res protein
structure vs. a low res? I had a feeling that
super low rmsd is partly abused as a sign of crystallographic
prowess, like building fantasy stuff into nonexistent density
just to have no missing residues &c &c &c...
Thx, br
Hi Bernhard,
What I wrote means that unless you have superhigh resolution (in which
case you can simply forget about library restraining --- you can
even count the number of electrons for one atom to assign its type) the
model should reflect reality.
You state that reality is RMSD = 0.014, I thought it was 0.012 A. Small
mistake from my part. Anyway the message was simply to state that
targetting an RMSD of 0.02 meant that the geometry was not restrained
tightly enough to fit what one expects at a resolution where unrestrained
refinement cannot be done (who would attempt unrestrained refinement at
3.4 A refinement?).
Fred.
--
s-mail: F.M.D. Vellieux (B.Sc., Ph.D.)
Institut de Biologie Structurale J.-P. Ebel CEA CNRS UJF
41 rue Jules Horowitz
38027 Grenoble Cedex 01
France
Tel: (+33) (0) 438789605
Fax: (+33) (0) 438785494
e-mail: [EMAIL PROTECTED]
--
****************************************
Dirk Kostrewa
Paul Scherrer Institut
Biomolecular Research, OFLC/110
CH-5232 Villigen PSI, Switzerland
Phone: +41-56-310-4722
Fax: +41-56-310-5288
E-mail: [EMAIL PROTECTED]
http://sb.web.psi.ch
****************************************
