Dear Matthias,

the idea of paired refinement is to quantify the amount of information in your 
structure. For that purpose it is most likely counterproductive to disturb 
your parameter at all! In the worst case you end up comparing different 
structures.

Use the very same structure, do not reset B-values, do not stir or shake, just 
refine until the target value stabilises. With refmac5 and shelxl this happens 
usually quite quickly, a few hundred cycles should be sufficient. 

Best regards,
Tim

On Sunday, August 20, 2017 5:06:42 PM CEST Matthias Barone wrote:
> Thanks James for your reply. I my naive way, I thought that when doing
> paired refinement, it would be sufficient to run several cycles of
> refinement with simulated annealing followed by a thorough real space
> rebuilding. What is your thought about using  the "jiggling" of additional
> cartesian annealing steps (phenix) to check the influence of, say, an
> additional 0.1A outermost shell? I used paired refinement over a couple of
> cycles and thoroughly deleted what was not visible anymore. Less for the
> removal of bias, but more to increase degrees of freedom... But now that I
> read your answer, I have the impression that this was not enough to remove
> bias of the sidechains?
> 
> matthias
> 
> >>> James Holton <[log in to unmask]> 08/19/17 7:23 PM >>>
> 
>      Yes, B factors are indeed a super-absorbent sponge for model bias.     
> Best to re-set those before re-refinement.  But you should also do    
> something with the coordinates, and probably the occupancies too.
> 
>      My jiffy to add to Graeme's list of options is this script:
>      http://bl831.als.lbl.gov/~jamesh/scripts/jigglepdb.awk
>      It has a variety of options, including separate rms shifts for    
> coordinates, B factors and occupancies (shift=, Bshift=, and     Oshift=),
> and a special option called "shift=byB" that moves     coordinates in
> accordance to the atom's B factor.  The distribution     of the shifts can
> be Gaussian (default), Lorentzian or uniform     sphere.  The latter is
> useful if you want to avoid very large shifts     (see below).  By default,
> one conformer (i.e. A, B, C) is randomly     given an occupancy of "1" for
> all residues and all the other     conformers are given  zero.  This is an
> attempt to simulate the     cell-to-cell variation of the structure, which
> can only have one     conformer at a time.  If you specify "keepocc=1" this
> behavior gets     turned off.  Any non-zero "Oshift" will add random noise
> to all     occupancies.
> 
>      I'm sure all this functionality would be easy to re-create using    
> CCTBX as well.
> 
> 
>      I've actually played around with "jiggling" pdb files a fair bit.     
> Turns out that in order to remove "bias" completely you need to kick    
> the atoms by an rms distance comparable to the relevant resolution.     
> That is, 1.5 A for 1.5 A spots, but also 6 A if you want to un-bias     6 A
> spots.  This tends to be outside the radius of convergence of     most
> refinement programs.  In fact, kicking atoms by as little as     0.5 A can
> often result in some side chains falling into alternate     rotamers and
> the like.  Waters can also fall out of place and drift     into nearby
> pockets of liberated density.  You can do a "cleanup"     after
> re-refinement to correct for such gross errors.  Looking for     the
> largest differences between starting and re-refined models     generally
> lights these up.  But before you "fix" these things you     have to start
> asking questions about what "bias" really is.
> 
>      It seems a popular "bias-reduction" approach is to re-run Phaser or    
> other molecular replacement job, but all that really does is put     your
> model onto an alternative origin or asu choice.  These choices     are
> arbitrary, of course, and crystallographically equivalent.  Once     you
> have corrected for the origin shift using something like     "reforigin"
> you can see that all your Phaser run has produced is a     slight
> rigid-body shift relative to where you started.  Hardly a     bias-removal
> technique.
> 
>      So, what is "bias"?  We tend to think of "bias" the same way we    
> think of "twinning": as a synonym for "evil".  Something to be kept     out
> of our models and data at all costs.  But if you think about it     a bit
> you may realize that even having the backbone structure itself     intact
> is a form of "bias".  As in you are "biasing" your model to     resemble
> the overall fold you originally assigned to the structure.      You may or
> may not be all that worried about this.  And if you are     certain that
> the main chain trace is correct, then enforcing it in     further analysis
> is not "bias", it is "prior knowledge".  I suppose     knowing the
> difference between these two things is what science is     all about.
> 
>      So, I'd say that "resetting" a structure depends on what aspect of    
> the structure you're trying to test.  If you made a mistake in the    
> backbone trace, or even a rotamer assignment, then doing a 0.5A     jiggle
> isn't going to reset that mistake.  But if your trying to     test the
> influence of data between 1.5 A and 1.4 A, then I'd say do     a jiggle of
> at least half that distance.
> 
>      -James Holton
>      MAD Scientist
> 
> 
>      On 8/17/2017 8:40 AM, Robbie Joosten       wrote:
> 
>                                  *.EmailQuote {
>     margin-left: 1.0pt;
>     padding-left: 4.0pt;
>     border-left: rgb(128,0,0) 2.0px solid;
> }
>               p.x_MsoNormal, li.x_MsoNormal, div.x_MsoNormal {
>     margin: 0.0cm;
>     font-size: 11.0pt;
>     font-family: Calibri , sans-serif;
> }
> a:x_link, span.x_MsoHyperlink {
>     color: blue;
>     text-decoration: underline;
> }
> a:x_visited, span.x_MsoHyperlinkFollowed {
>     color: rgb(149,79,114);
>     text-decoration: underline;
> }
> *.x_MsoChpDefault {
> }
> div.x_WordSection1 {
> }
>                            In most cases resetting the B-factors            
> would be enough to perturb the model.
> 
>            Cheers,
>            Robbie
> 
>            Sent from my Windows 10 phone
> 
>                         From:               Andrew Leslie
>                Sent: 17 August 2017 17:29
>                To: [log in to unmask]
>                Subject: Re: [ccp4bb] "reset" a structure before             
>  re-refinement
> 
> 
> 
> 
>                   Hi Graeme,
> 
>              You can do this with PDBSET, keyword NOISE
>              Cheers,
> 
> 
>              Andrew
> 
>              > On 17 Aug 2017, at 16:17, Graeme Winter            
>              > <[log in to unmask]> wrote:
>              > 
>              > Dear All,
>              > 
>              > Is there a protocol out there to gently perturb atomic       
>              >      positions so that re-running refinement can essentially
>              > put             them back without bias from the original
>              > refinement? In             particular, if trying to perform
>              > the Karplus and Diederichs             paired refinement
>              > protocol, I do not want to run the lower            
>              > resolution refinements with the "memory" of the weak high   
>              >          resolution data present... and only have the
>              > refined             structure to work from...
>              > 
>              > Am using refmac5, but any pdb randomizer would hit the       
>              >      spot
>              > 
>              > Many thanks Graeme

-- 
--
Paul Scherrer Institut
Tim Gruene
- persoenlich -
OFLC/104
CH-5232 Villigen PSI
phone: +41 (0)56 310 5297

GPG Key ID = A46BEE1A