Re: Post-modeller refinement

[Andrej Sali <sali@rockvax>]

Hi,

When the input alignment is correct, it is very rare to find
stereochemical 
violations in models built by the default modeller optimization,
especially when more than one model is built (say 20) and the best one
is evaluated. Violations usually happen in those regions of a model that
have mutually inconsistent restraints applied to them (eg, because of an
incorrect alignment (especially when more than one template is used),
because of some other user mistake, or a bug in the program). Also, how
large are the violations - are they really serious (eg, > 5 standard
deviations)? 

If one is not certain about the alignment in the problematic region, I
suggest changing it to other reasonable possibilities, especially if
more than one template is used. 

MD refinement, in solvent or in vacuo, in general increases the RMS
error of a model (see the F. Cohen et al. review of CASP2 in Folding and
Design, 1997; or the Janet Thornton review in the CASP2 Proteins
supplement 1998). This of course does not mean that it is not useful to
do sometimes, ;-)

Best, Andrej


Nick Glover wrote:
> 
> Modellers:
> 
> I have built a model, and submitted it to what I believe is a fairly
> rigorous optimization schedule (comments?):
> 
> SET LIBRARY_SCHEDULE = 1
> SET MAX_VAR_ITERATIONS = 500
> SET MD_LEVEL = 'refine1'
> SET REFINE_HOT_ONLY = 0
> SET RSTRS_REFINED = 2
> SET REPEAT_OPTIMIZATION = 5
> 
> However, having checked the model with the Biotech Collaboration on-line
> validation tools I note several errors in my structure, notably in terms
> of VDW overlaps, bond angles, and unfulfilled buried h-bond formation,
> consistent with ENERGY evaluation which indicates a number of 'serious
> non-bonded atom clashes' and other violations.
> 
> Ramachandran plots confirm the good stereochemistry of the model.
> 
> Now, looking through Dr Sali's various papers I note that he performs no
> post-model refinement.. however, I wonder whether the community would
> care to comment on their own experiences in terms of model evaluation,
> and whether they have performed post-minimization or MD to relieve poor
> contacts, etc.
> 
> My suggestion: in-vacuo steepest descents minimization followed by
> conjugate gradients to less than 0.1 kcal/ang., performed with a
> restraining potential on the backbone to preserve the backbone fold of
> the model, while allowing the sidechains to re-orient and relieve any
> bad contacts..
> 
> obviously, this is a fairly tame refinement, and could be performed in a
> solvated shell, etc.
> 
> Suggestions and comments appreciated... a few words from those who have
> built and evaluated models would be very beneficial to those new to this
> methodology!
> 
> Nick
> 
> ----------------
> 
> NB: for interested parties..
> 
> The validation suite (Procheck, what-if, etc,) is available for on-line
> model verification:
> 
> http://biotech.embl-heidelberg.de:8400
> 
> ----------
> 
> Nick Glover
> Department of Chemistry
> Simon Fraser University
> 8888 University Drive
> Burnaby, BC
> V5A 1S6
> Canada

--
Andrej Sali, Assistant Professor
The Rockefeller University, 1230 York Avenue, New York, NY 10021-6399
tel +1 212 327 7550; lab +1 212 327 7206 ; fax +1 212 327 7540
e-mail ; http://salilab.org