[phenixbb] Weighting in phenix.real_space_refine

[Pavel Afonine]

Hi Oliver,

the refinement target is

T = weight * Tdata + Trestraints .

So bigger weight makes map contribution more important and smaller 
weight emphasizes restraints.

In your example, most segments yield weight=10 with a few outliers being 
less than 1. The average weight used in refinement is about 10 (computed 
after filtering outliers).

Technically it is possible to implement finding optimal weight for each 
segment but as you can imagine this is going to be terribly slow.

I have not experimented with "optimal weight vs local map resolution". 
But *speculating* I would say this is not important because refinement 
target used in phenix.real_space_refine does not depend on map shape but 
only map peaks.

Of course you can try splitting domains and refining them separately and 
then assembling the full model back. But the differences you may see may 
not necessarily be due to the weight (they may be because domains don't 
see neighbors, for example).

It would be very nice to identity a very specific and clear example (and 
small too so refinements run fast!) where using local weights vs global 
actually does make a difference.. May be you or someone on the list has 
something like this? That would be a starting point for me to do some 
research in this direction.

Thanks for pointing “Dist moved from previous” issue - I will take care 
of it..

Pavel

On 9/13/16 08:12, Oliver Clarke wrote:
> Hi Pavel,
> So how do I interpret the weights in the output? For example, with a structure I am refining at the moment, I have the following:
>
> "Weight determination summary:
>     number of chunks: 40
>     random chunks:
>     chunk 33 optimal weight:    0.9990
>     chunk 30 optimal weight:   10.0000
>     chunk 16 optimal weight:    9.9990
>     chunk 10 optimal weight:    9.9990
>     chunk 20 optimal weight:    9.9990
>     chunk 16 optimal weight:    9.9990
>     chunk 31 optimal weight:    0.0040
>     chunk 12 optimal weight:    5.9990
>     chunk 19 optimal weight:   10.0000
>     chunk 23 optimal weight:    0.8980
>     overall best weight:    9.4279”
>
> Is low tight and high loose, as for phenix.refine? And are there any prospects for the future of varying weights across the structure? Often in EM we will have  large regions at good resolution, say <4Å, and other regions which while well defined are at lower resolution, say 6Å - applying the same weights to both regions doesn’t seem ideal.
>
> Would your suggested strategy at the moment be to split the structure domain-by-domain, refine domains separately and then recombine?
>
> Cheers,
> Oli
>
> (Also, there seems to be a very minor bug in the output model stats - “Dist moved from start” changes, but “Dist moved from previous” is always 0.000.
>
> Model statistics:
>     Map CC (whole unit cell):  0.662
>     Map CC (around atoms):     0.773
>     rmsd (bonds):              0.01
>     rmsd (angles):             1.04
>     Dist. moved from start:    0.060
>     Dist. moved from previous: 0.000
>     All-atom clashscore        4.47
>     Ramachandran plot:
>       outliers:                0.27  %
>       allowed:                 12.24 %
>       favored:                 87.49 %
>     Rotamer outliers:          0.54 %
>     C-beta deviations:         4)
>
>> On Sep 13, 2016, at 11:01 AM, Pavel Afonine <pafonine at lbl.gov> wrote:
>>
>> Hi Oliver,
>>
>>> So just to be sure - the average weight obtained with the randomly selected fragments is used for the entire model?
>> correct.
>>
>> Pavel
>>