Brestraints, G. Kleywegt's comments

Manfred Weiss msweiss at imb-jena.de
Thu Jun 6 11:18:50 EST 1996


In the discussion about temperature factor restraints
G. Kleywegt wrote:

> 
> * B-factor restraints test @ 960604
> ***********************************
> 
> * compare default B restraints in X-PLOR with Dale's restraints
>   as posted on the xplor newsgroup
>   (NOTE: this post does not quite implement Dale's method:
>   - XPLOR would need something like: bsigma=(selection)=(selection)=target
>     as it is now, every atom gets the target of the last bsigma
>     statement it fits; for example:
>  bsigma = ( name N  or name CA ) = 4.6
>  bsigma = ( name CA or name C  ) = 7.4
>     -> CA will get a target of 7.4 anyway
>   - Dale's set has *directionality*)
>   (NOTE: critique on Dale's paper: he never used Rfree to show that you
>    actually get better models at any resolution !)
> 
> * first try with CRABPII holo, 1.8 A data, use near-final model
>   still with Rfree (i.e., prior to final refinement against all data)
>   do the following for several resolutions:
>   - set all bs to 15 + Gaussian term as per normal xplor
>   - refine with default xplor restraints
>   - set bs to 15 + gaussian again
>   - refine with pseudo-Dale restraints
> 
>   Resol		R/Rfree start	Normal Bs	Dale's Bs
>   2.0		0.270/0.287	0.202/0.226	0.193/0.220
>   2.2		0.259/0.275	0.197/0.220	0.185/0.216
>   2.4		0.250/0.267	0.191/0.218	0.178/0.212
>   2.6		0.242/0.263	0.184/0.219	0.169/0.214
>   3.0		0.230/0.252	0.175/0.211	0.161/0.205
> 
>   -> in all cases pseudo-Dale is ~0.5 % better in Rfree
>   however, also R is lower, but usually 2-3 times as much as Rfree
>   -> pseudo-Dale seems to be slightly more accurate, but
>   it also increases the extent of over-fitting  (since R drops
>   quite a bit more than Rfree)
> 
> * try the same with CBH2 native data (1.8 A):
> 
>   Resol		R/Rfree start	Normal Bs	Dale's Bs
>   2.0		0.232/0.247	0.169/0.198	0.159/0.197
>   2.4		0.215/0.232	0.156/0.193	0.143/0.191
>   3.0		0.196/0.210	0.140/0.182	0.123/0.179
> 
>   -> same picture: 0.1-0.3% drop in Rfree at the expense of
>      a drop in R which is 5-10 * larger
> 
> * -> stick to the old restraints in X-PLOR
> 
> 
> --------------------------------------------------------
> Gerard CD Kleywegt
> Presently at Yale (mailto:gkleyweg at laplace.csb.yale.edu)
> Normally in Uppsala (mailto:gerard at xray.bmc.uu.se)
> 

I basically agree that there is a possibility of overfitting
with the 'pseudo-Dale' restraints. But there also seems to
be some information in them. So why stick to the old ones
if we can do better.
I propose a better implementation of the idea, rather than
just stick to whatever we did before.
But then, who's got the time?

  Manfred Weiss

-- 

   Manfred S. Weiss, Dr.
   Institute of Molecular Biotechnology
   Department of Structural Biology and Crystallography
   Beutenbergstrasse 11         
   Postfach 100813                                   Phone: +49-3641-65-6070
   D-07708 Jena                                      Fax:   +49-3641-65-6062
   GERMANY                                           Email: msweiss at imb-jena.de




More information about the X-plor mailing list