[phenixbb] Twin refinement issues?

[Phil Jeffrey]

Hi Peter,

Thanks for the reply.

So not a bug, just a Really Undesirable Feature.

I'm familiar with point group 32 and actually ran the program with both 
indexings, which unexpectedly came up with roughly the same R-free 
(irony: the incorrect indexing was marginally better).  Mea culpa for 
not using my standard m.o. of comparing datasets using CCP4's SCALEIT.

The reasons I consider this behavior to be a bad thing are:

1.  Twin fractions > 0.5 are not physically reasonable and I would 
expect default behavior should be to clamp them to the range 0.-0.5. 
(caveat: I'm not sure what CNS or REFMAC do).

2.  I have mixed feelings about "stealth reindexings", especially since 
I didn't set refinement.second_guess=True  (see below)

3.  The stealth reindexing behavior is inconsistent across refinement 
without twinning set (no stealth indexing) and when it is set, for 
exactly the same mtz and pdb files.

Auto-reindexing might be a time-saving feature (especially in point 
group 3) but should be consistent w/ and w/o twinning enabled and needs 
to be much more prominently advertised in the output, and a new MTZ file 
written.

Aside: this structure was the result of a MAD dataset at 3.5 Angstrom 
with a lower twin fraction (~0.2) so of course the relative indexing 
problem comes into play when I switched datasets.  In the case of the 
MAD dataset SHARP made a radical difference in phase quality compared to 
  the uninterpretable map straight out of SHELX - perhaps the largest 
improvement I have ever seen.

Cheers
Phil Jeffrey

Peter Zwart wrote:
> Hi Phil,
> 
> twinning with a twin fraction of 1-alpha is the same as twinning by
> alpha and reindexing your data (with the twin law for instance).
> Not a bug, but a consequence of structure solution, indexing
> ambiguities. Did you run MR, or did you have 'old' model already?
> 
> 0.59 is the same as 0.41 (after reindexing), pretty close to the
> britton and H test (0.35). The ML test typically is lower then the
> other two estimates, as it tries to incorporate experimental errors.
> 
> HTH
> 
> Peter
> 
> 
> 
> 2009/4/9 Phil Jeffrey <pjeffrey at princeton.edu>:
>> While experimenting with twin refinement for a crystal in P321 that
>> approximates P622 to varying extents, I noticed:
>>
>> | twin fraction: 0.59  twin operator: -h,-k,l
>>       |
>>
>> A bug, perhaps ?
>>
>> This twin fraction would seem to be unlikely, since twin fractions > 0.5
>> have no meaning.  Since this data does not scale well in p622 compared
>> to some more heavily-twinned datasets on this xtal form I think the twin
>> fraction is certainly less than 0.5.
>>
>> Xtriage estimates:
>> Statistics depending on twin laws
>> -----------------------------------------------------------------
>> | Operator | type | R obs. | Britton alpha | H alpha | ML alpha |
>> -----------------------------------------------------------------
>> | -h,-k,l  |   M  | 0.150  | 0.343         | 0.347   | 0.283    |
>> -----------------------------------------------------------------
>>
>>
>> Unless of course phenix.refine is reporting 2*alpha.
>>
>> Phenix v1.4-3  Intel Mac OSX 10.5.6
>>
>> This particular run with:
>>
>> phenix.refine model-08.pdb pz7e_truncate-unique.mtz
>> refinement.main.ncs=true strategy=individual_sites+group_adp --overwrite
>> xray_data.r_free_flags.generate=True twin_law="-h,-k,l"
>>
>> This is a very early non-finessed model.
>>
>> Phil Jeffrey
>> Princeton
>> _______________________________________________
>> phenixbb mailing list
>> phenixbb at phenix-online.org
>> http://www.phenix-online.org/mailman/listinfo/phenixbb
>>
> 
> 
>