Re: refmac for partial p-ala trace

["Eleanor J. Dodson" <ccp4@ysbl.york.ac.uk>]

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

kostrewa wrote:
> 
> ***  For details on how to be removed from this list visit the  ***
> ***    CCP4 home page http://www.dl.ac.uk/CCP/CCP4/main.html    ***
> 
> Anastassis Perrakis wrote:
> 
> > What really matters is not the resolution of the map (ie phases) that you build the model on,
> > but really the quality of the native dataset, or any relevant native dataset to be exact
> > (ie another crystal form).
> >
> > As soon as you have a crude (but correct) model you are only limited by the native data ...
> >
> > I have met quite a few people out there, that still do the 'good-old-protocoles', first refine
> > to 2.8, then to 2.5, then to 2.3, etc etc. Maybe it is time to emphasize again
> > that with using exp. sigmas and max.lik., strategies like that are obsolete ?
> 
> Hmm, at least in my experience, the real strength of maximum-likelihood refinement comes into play when you have observed phases with reasonable figure-of-merits. Without that, ML refinement also suffers from phases that are too heavily biased towards the (partial) model phases. Of course, with an almost complete and correct model, the maximum resolution of the best data set is
> the limiting factor for the quality of the refined model. However, for partial errornous models without reasonable observed phases, I think one still has to be cautious. Refinement against increasing resolutions was a successful protocol for least-squares methods. Theoretically, with ML, the high resolution terms should be appropriately down-weighted to "automatically" follow
> such a protocol. I think, currently only BUSTER does that effectively. So, trying one of those 'good-old-protocols' with REFMAC or CNS might still be a good idea. As Tassos said, reducing the number of parameters by torsion angle dynamics or application of any reasonable density modification might also help to reduce model bias. The reliability of experimental sigmas is another
> hot topic.
> 
> Good luck,
> Dirk.
> 


  Well - we find ML refinement works quite well with poor or incomplete
MR models, with or without observed phases. ( They help a lot PROVIDING
the FOMs are sensibly estimated..) 

And including all the data works better than restricting it, providing
you are estimating errrors from the Free R set. Initially the weights
assigned to the outer resolution bins are very small, but they still
contribute significantly to the important overall scaling parameters. 

The amount of bias is often a function of the amount of missing data. By
default REFMAC substitutes these reflections as Dfc instead of the
2mFo-dFc used for observed data, and any such term will introduce bias..




But there is a limit! 15% of the atoms seems to be  too little; 50% is
more than enough . There must be a cut off somewhere in between..
Eleanor Dodson