Re: [ccp4bb]: ESD of distances!

[<rams@poori.biochem.uiowa.edu>]

***  For details on how to be removed from this list visit the  ***
***          CCP4 home page http://www.ccp4.ac.uk         ***


Hi Ethan,

I wanted to read the paper in JMB before I reply.  I agree with you on the 
underestimation of ESD's.  I dont follow the Cowtan and TenEyck paper 
enough to comment. However, I disagree that calculating ESD's with the 
BLOC card is not useful.  It is better than not getting ESD's at all.
However, one must be certain that one mentions that the ESD's are not 
accurate and do not reflect or even come close to what one gets.  
IF the same protein is refined the same way in two different complexes and 
one finds a difference (especially in non-bonded distances), how does one 
know the difference is genuine?  Say, if the diffence is 0.4A. Now, even 
with the biased ESD's, if one gets that the accuracy of the measurement 
of length is 0.03A, this is a significant difference.  This kind of a 
quasi-quantitative number is assuring (falsely comforting ?) than  the 
absence of it. 

Yes, I will have to do the complete eigen value analysis as suggested by 
Cowtan and TenEyck to know if this is biassed - I agree.  

Now, I would like to calculate the full matrix least squares.  Why does 
the compiler complain if I increase the size of the array B ( in SHELX), 
bigger than a given number in FORTRAN.  Given 64 bit machines, is there 
a  version of SHELX where I can do 800,000 reflections and 120,000 
parameters with full L.S. IF somebody has tweaked SHELX to do this, I 
would love to use it.

Thanks for all the replies. 

Rams.



On Mon, 18 Mar 2002, Ethan Merritt wrote:

> ***  For details on how to be removed from this list visit the  ***
> ***          CCP4 home page http://www.ccp4.ac.uk         ***
> 
> I had snipped the part where he said that he had used shelx to calculate
> esds.  That option only exists in L.S. mode, not in CGLS.
> 
> It is possible to use L.S. refinement in block mode (BLOC) rather than
> true full-matrix, and ask for esds.   May I recommend to people that you
> not do this, however.  Or at least if you do, be aware that the esds 
> estimated from blocked matrix treatment are under-estimates of the
> "true" esds you would get from full-matrix treatment.   This is because
> the covariance of parameters in separate blocks is necessarily omitted
> from the calculation.  If you partition into, say, 5 blocks then you are only
> sampling 1/5 of the covariance terms in the full matrix.
> 
> I have found for a ~500 residue protein that partitioning the calculation
> into 5 blocks of 100 residues (plus a 6th for  bound oligosaccharide and 
> solvent molecules) causes the esds to be underestimated
> by roughly 10% relative to the values obtained from true full-matrix.
> [Merritt et al (1998) J. Mol. Biol. 282, 1043-1049].  I have subsequently
> obtained similar numbers using other structure refinements.
> For a more mathematical investigation of how covariance distributes itself in
> the matrix you might also have a look at the paper by Kevin Cowtan
> and Lynn Ten Eyck  in Acta  D 56, 842-856 (2000).
> 
> Is 10% underestimate significant?  [shrug] Well, I assume that if you are
> going to the trouble of calculating esds then you care enough to want 
> accurate values.  And the amount of underestimation will in any case be
> dependent on the actual distribution of covariance in your particular
> matrix, so I cannot even predict how typical  my  value of 10% would
> turn out to be.
> 
> On a related note - I have written a Perl script to extract the esds reported
> in a shelx output log and reconstruct a PDB file containing the appropriate
> SIGATM records.  I'd say it's in beta-test condition, if anyone wants me to
> Email a copy for evaluation. 
> 
> 
> 

-- 
S. Ramaswamy
Department of Biochemistry.
University of Iowa.