Dear Tim,
There are low resolution structures and low resolution structures. At one extreme, the low resolution structure might have a 0.3Å rmsd with a 0.9Å native structure, and at the other extreme, the data might be twinned, there might only be a remote, incomplete model and the electron density maps may look really bad.
In my experience, the programs I work with (Buster, Refmac) do not always produce the optimal weights at low resolution in automatic mode and I fully agree with Bernard that at low resolution the geometry needs to be tightly restrained to prevent model bias.
One could of course activate the button to "Execute multiple refinement runs". However, from just pressing buttons one does not learn crystallography and one does not get a feeling how reliable and good or bad a structure really is.
My 2 cts,
Herman
-----Ursprüngliche Nachricht-----
Von: CCP4 bulletin board [mailto:[log in to unmask]] Im Auftrag von Tim Gruene
Gesendet: Mittwoch, 27. Januar 2016 14:59
An: [log in to unmask]
Betreff: Re: [ccp4bb] Refmac auto weight
Dear Bernhard,
I doubt that phenix and Refmac5 produce the same results only because the R- values are similar - and you surely know this!
I tend to integrate data myself and when it seems sound and the integration is stable, I would go with the auto weighting scheme in Refmac5. If the geometry (RMS values) deviates from what I am used to, I start looking at manual weight fitting. At medium to low resolution I would use quite a low weight, 0.02 or less, and increase the number of cycles to e.g. 100 or more.
Did you take a look at the -LL plot vs. cycle to monitor if the refinement converged?
In ccp4i2 there is an Advanced Option to "Execute multiple refinement runs in order to optimise X-ray/geometry weight". How do the figures look like when you activate this button?
Best,
Tim
On Wednesday, January 27, 2016 12:15:30 PM Bernhard Rupp wrote:
> Dear Developers & Friends,
>
>
>
> Win CCP4i 6.5.019
>
> Win Refmac_5.8.0135 version 5.8.0135
>
>
>
> I have noticed (as have many others) that occasionally the Refmac
> auto-weight of restraints does not
>
> provide optimal results. I have a simple case (Fab, 1.9A, SG19) where
> significantly better stats can be obtained by
>
> the usual hand-optimization, with the end result Rs and rmsds being
> almost the same as the auto-weighting
>
> obtained by Phenix.
>
>
>
> So, does any one of the high-volume refiners like Robbie have any
> indications under
>
> what circumstances the refmac procedure is suboptimal? What exactly is
> the refmac procedure? And if
>
> one can do it by hand and phenix automatically gives the same results,
> could it be implemented in a similar
>
> fashion in Refmac?
>
>
>
> Same coordinate input file
>
>
>
> Auto : initial final
>
> R factor 0.1835 0.1828
>
> R free 0.2415 0.2419
>
> Rms BondLength 0.0158 0.0157
>
> Rms BondAngle 1.9467 1.9511
>
>
>
> Manual, 0.05:
>
> R factor 0.1835 0.1957
>
> R free 0.2415 0.2357
>
> Rms BondLength 0.0158 0.0070
>
> Rms BondAngle 1.9467 1.4095
>
>
>
> Phenix auto:
>
> R factor 0.1902 0.1992
>
> R free 0.2444 0.2387
>
> Rms BondLength 0.0160 0.0090
>
> Rms BondAngle 1.8180 1.1795
>
>
>
> In any case, I think aiming for small molecule rmsds regardless of
>
> resolution and thus data/params is not ideal. Upper limit, yes. Many
>
> others have voiced the same, and George Sheldrick once showed a
>
> nice graph of restraint weight (or rmsd?) vs resolution as obtained by
> his
>
> 'likelihood by intuition' with a sigmoid curve showing the general
> trend
>
> that lower res needs higher restraint weights. Actually, showing more
> 'restraint'
>
> when modelling at low resolution seems a good idea in general (just
>
> to add something vitriolic.)
>
>
>
> Best, BR
--
--
Paul Scherrer Institut
Dr. Tim Gruene
- persoenlich -
OFLC/102
CH-5232 Villigen PSI
phone: +41 (0)56 310 5297
GPG Key ID = A46BEE1A
|