RE: [ccp4bb]: Refmac vs. cns

[Kostrewa Dirk <dirk.kostrewa@psi.ch>]

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

Okay, let's jump into the fire.
Initially, I was very enthusiastic when simulated annealing was introduced
in macromolecular refinement. My enthusiasm quickly turned into skepticism,
when in several projects with resolutions ranging from 1.7 to 2.5 Angstrom I
looked both at the resulting structures and the electron density maps: I had
to go through the whole structure again to correct all these little errors
(side chain and main chain) that simulated annealing obviously has
introduced; the electron density maps showed a lot of model bias, i.e.
too-good-to-be-true density around the "refined" model, and virtually flat
density for the missing parts. Thus, simulated annealing heavily over-fitted
my structures, which was presumably the reason why Axel Bruenger had to
introduce the concept of the Free-R factor and the use of multiple simulated
annealing rounds with averaged electron density maps. I also tried the
combination of torsion angle dynamics with maximum likelihood target, and it
seems to give good composite-omit-maps, but apart from that, I don't see any
real advantage over careful inspection of electron density (omit) maps and
fitting by hand. Dave Banner thinks, that it might still be very useful when
the resolution is low (say, around 3 Angstrom), and in your model is a
beta-strand with many short side chains (Ser, Thr, Val, Asx) that have to
flip around simultaneously along with a slight shift in the main chain. But
at that resolution, model bias is a very serious problem and it will be
difficult to judge the results. At high resolution, simulated annealing
should work very well, but there are more efficient algorithms available
(ARP/wARP, for instance).
To summarize, my personal verdict is that simulated annealing in
macromolecular refinement is a heavily overestimated technique.

So, I should better switch off my e-mail program now.

Dirk Kostrewa.

". . . and he walked a little further 
and he found he was alone 
and the wind, it gathered round him 
he was thinking of his sins 
he was looking at his song-bird 
and he was looking at his wings 
and there was trouble taking place"
P. J. Harvey

*****************************************************
Dirk Kostrewa
Paul Scherrer Institut   e-mail: dirk.kostrewa@psi.ch
Life Sciences             phone: +41-56-310-4722
OSRA/007                    fax: +41-56-310-4556
CH-5232 Villigen PSI        WWW: http://sb.web.psi.ch
Switzerland
*****************************************************


> -----Original Message-----
> From: Eleanor J. Dodson [mailto:ccp4@ysbl.york.ac.uk]
> Sent: Montag, 21. Januar 2002 10:12
> To: Jianghai Zhu
> Cc: ccp4bb@dl.ac.uk
> Subject: Re: [ccp4bb]: Refmac vs. cns
> 
> 
> ***  For details on how to be removed from this list visit the  ***
> ***          CCP4 home page http://www.ccp4.ac.uk         ***
> 
> Jianghai Zhu wrote:
> > 
> > ***  For details on how to be removed from this list visit the  ***
> > ***          CCP4 home page http://www.ccp4.ac.uk         ***
> > 
> > Hi,
> > 
> > I am very interested in this comparison too. Can anybody 
> give me some
> > details about the comparison between Refmac5 and CNS? I 
> have never use
> > Refmac5 before. Can refmac5 do simulated annealing?
> > 
> > Thanks.
> > 
> > ================================
> > Jianghai Zhu
> > Biochemistry & Molecular Biology
> > B203, Lilly Hall
> > Purdue University
> > Tel: 765-4949247 (O)
> >       765-4630228 (H)
> > ================================
> 
> 
>  Heres a fire storm brewing..
>  Can some-one comment on how valuable simulated annealing is 
> at various
> resolutions?
> 
>  In examples at better than 2.5A we usually seem to finish up 
> doing just
> about the same amount of rebuilding as you do after a ML run. 
> 
>  Eleanor
>