Re: [ccp4bb]: Summary: Amore Rotation Function Scoring?

[<rjr27@cam.ac.uk>]

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

> > "The Amore cross-rotation function basically calculates a correlation
> > coefficient between the observed and calculated Patterson function (CC_P).
> > However, the output of the cross-rotation search is for some dubious reason
> > sorted on the correlation coefficient between calculated and observed F
> > (CC_F). This doesn't make much sense to me for the following reasons:
> > (1) The search function is the CC_P, thus, from a methodological point of
> > view, the output should be sorted on this value and not on something else.

I don't really agree with the logic of this.  The fast rotation function
takes the form it does because this can be computed quickly.  The
philosophy of AMoRe is to use a fast scoring scheme to generate plausible
solutions quickly, but then to re-evaluate them based on a better score.
There's no overriding reason for that score to be based on Patterson
overlap like the fast function.

> > (2) Both, the calculated F and I of the model only make sense after it has
> > been correctly positioned, which is not the case in the cross-rotation
> > search.

This is also not true.  As we showed in the paper on BRUTE, a correlation
coefficient on intensities is equivalent to the correlation coefficient
between the corresponding origin-removed Patterson maps.  Because the
origin-removed Patterson map includes the self vectors from which the
orientation can be judged, the calculated I (of a single oriented model
in a P1 cell) does make sense before it is correctly positioned.

> > (3) Accordingly, the signal-to-noise must be much better for CC_P than for
> > either CC_F or CC_I. To illustrate this, I have run a cross-rotation search
> > with the refined protein-only model of the A. niger phytase (Kostrewa et al.,

Perhaps Jorge is reading this and will comment, but my recollection of
what he's said in talks is that he has chosen the criterion on which he
sorts by running tests on a wide variety of cases.  In an individual
problem it may not give the best results.

Now a bit of a plug.  If a molecular replacement problem is difficult
enough that the different criteria in AMoRe give different choices of
solution, then it's probably worth running Beast, because the
likelihood-based score really does seem to discriminate better.  You
can even just rescore the top solutions output from AMoRe or Molrep
to get them resorted by likelihood score.  (Note that, if you're using
AMoRe you have to be a bit careful, and use the reoriented/repositioned
model to which the AMoRe results refer.)

-- 

Randy J. Read
Department of Haematology, University of Cambridge
Cambridge Institute for Medical Research      Tel: + 44 1223 336500
Wellcome Trust/MRC Building                   Fax: + 44 1223 336827
Hills Road                                    E-mail: rjr27@cam.ac.uk
Cambridge CB2 2XY, U.K.                       www-structmed.cimr.cam.ac.uk