On Thu, 8 Apr 1999, Alexandre Alechine wrote: > I've found that omitting poorly resolved side-chain atoms > from refinement often improves > R-free (sometimes up to 0.5%). The reason is trivial: > These side chains are on a surface of protein > and are naturally disordered. Because the B-factors of side and main chains > are mutually restrained the refinement produces wrong B-factor values. > Releasing restrains improves R_free too, but not as well as manual omitting > bad side chains. This is a problem which has been discussed by a number of people including Dale Tronrud and myself at various CCP4 Study Weekends and Refinement Workshops. The basic problem is that most refinement programs (X-plor I know for sure, but no doubt others as well) do not use a sensible weighting scheme for the thermal parameter restraints. I did an analysis of differences between unrestrained isotropic thermal parameters of bonded atoms in an atomic resolution (1A) polypeptide structure with good data, and obtained the following relationship for the contribution to the expected error in the Uiso's: sigma(U) =~ U^2. If you prefer it in terms of B's it's sigma(B) =~ B^2/8pi^2 (Formulae always seem to be simpler when you use U's!) This error arises partly from the fact that atoms with large thermal parameters obviously do not make a significant contribution to a large fraction of the data at high resolution, and partly from the isotropic approximation, i.e. atoms with large vibration amplitudes tend to be highly anisotropic. It is only the along-bond component of the displacement that can be restrained, but for highly anisotropic atoms this component bears little relationship to the spherically averaged thermal parameter. For B = 10A^2 the formula gives sigma(B) = 1.25; for B = 20, sigma(B) = 5; B = 40, sigma(B) = 20; B = 80, sigma(B) = 80. Most people seem to use a weighting scheme corresponding to sigma(B) = constant, where the constant is typically ~ 1 to 5A^2. This is tolerable for "typical" main-chain atoms with B = 10 to 20A^2, but clearly not for side-chains. The effect of applying thermal parameter restraints with unrealistically high weights is obviously to drag down the large parameter values, as you have already pointed out. However I would maintain that the solution is not to leave them out, but to weight them properly! If omitting "bad" side-chains improves Rfree that must mean you didn't have a correct model for them in the first place. -- Ian _______________________________________________________________________________ Ian J. Tickle | Phone: (+44) 171 631 6854 Department of Crystallography | Fax: (+44) 171 631 6803 Birkbeck College (University of London) | Email: i.tickle@bbk.ac.uk Malet Street, LONDON WC1E 7HX, UK | WWW: http://www.cryst.bbk.ac.uk/
This archive was generated by hypermail 2b29 : Tue Mar 20 2007 - 03:13:26 CST