Just to add my two ha'porth.

I discussed this some years ago with Garib, just after I'd added the  
anisotropic cutoff to the resolution limits in Mosflm (mentioned by  
Andrew below); as I remember (and this is an invitation to Garib to  
contribute and correct me here!), the answer went along the lines that  
the systematically missing volumes of data would affect the maximum  
likelihood model used (which used an isotropic resolution diffraction  
model?), so the refinement of a molecular model produced with  
anisotropic data would be affected badly.

I think I would tend to follow Pierre's route, though, since the model  
I implemented only defines the resolution along a* b* and c*, and the  
sigma cutoff is probably more flexible.

Having said that, there are one or two users out there who have used  
the anisotropic cutoffs and have been very happy with the results (or  
at least that's what they said to me).

On 16 Sep 2009, at 11:13, Pierre Rizkallah wrote:

> Hi Everyone,
>
> I echo Andrew's thanks for the summary offered by Justin.
>
> I would like to mention another way to trim anisotropic diffraction  
> patterns of the weak patches 'at source', as it were, in MOSFLM, by  
> specifying a sigma cut off applied to each image.
>
> from the manual:
> RESOLUTION [ <lowres> ] < highres> > [CUTOFF <sigcut>]
> The CUTOFF subkeyword allows the resolution limit of reflections  
> written to the MTZ file to be different for each image. The  
> resolution limit is set as that resolution at which I/sigma(I) drops  
> to below "sigcut". The I/sigma(I) for fully recorded reflections (if  
> any) is used, otherwise partials. Default sigcut 0.0
>
> I used this option in the past 2 years on one particular data set,  
> which certainly lost all its weak hi-res spots, leaving an ellipsoid  
> of reflections. The merging stats improved significantly, but the  
> completness in the outer shells went down significantly too. I did  
> not compare maps with and without cut-off, in this particular case.
>
> If the drop-off is severe, one can imagine the effect on the maps  
> would be streakiness. There is no win-win situation, unfortunately.  
> Some people set the resolution limit to be half-way between that of  
> the best and worst directions, and take everything, weak and strong.  
> It is a sacrifice of some good data for the sake of the overall  
> quality, and produces distortions of its own, maybe without the  
> streaks.
>
> Pierre Rizkallah
>
>
> **********************************************************************
> Dr. Pierre Rizkallah, Senior Lecturer in Structural Biology, WHRI,  
> School of Medicine, Academic Avenue, Heath Park, Cardiff CF14 4XN
> email: [log in to unmask]     phone + 44 29 2074 2248
>>>> A Leslie <[log in to unmask]> 16/09/09 9:24 AM >>>
> I would like to thank Justin for his summary of this topic, which I'm
> sure many people found of interest, and is very much in the spirit of
> the bulletin board.
>
> I would just like to correct one factual error, in that it has been
> possible to specify anisotropic resolution limits to MOSFLM for many
> years, the appropriate keywords (described in the MOSFLM "Help"
> documentation) are:
>
> RESOLUTION ANISO 3.5 2.5 2.5
>
> where the three values are the resolution limits along (or close to)
> a*, b*, c*.
>
> Unfortunately this option is not yet available in imosflm.
>
> I have not personally used this option and so cannot compare its
> efficacy relative to integrating isotropically and then applying an
> anisotropic limit such as Justin describes.
>
>
> Andrew Leslie
>
> On 15 Sep 2009, at 21:48, Justin Hall wrote:
>
>> Dear All;
>>
>> In response to my "Anisotropic Diffraction In Refinement", which
>> asked for suggestions for how best to proceed with refinement with
>> an anisotropic data set, I received a large number of responses
>> which overwhelmingly suggested using the UCLA Anisotropy Server (<http://www.doe-mbi.ucla.edu/~sawaya/anisoscale/
>>> ).
>>
>> The Anisotripy Server treats scaled/truncated data sets (I used
>> Scala and the old Truncate program). Fo and SigFo are analyzed with
>> respect to resolution in three dimensions and the data treated in
>> three steps:
>> 1) An elliptical resolution boundary is determined and applied.
>> 2) A purely anisotropic B-factor is applied to the Fo and SigFo data
>> to cause the data in all directions to fall off equally.
>> 3) A negative isotropic B-factor is then applied to the structure
>> factors to force the fall-off in the strongest direction to match
>> that of the original data, effectively meaning that the data are not
>> scaled to the mean but the weaker data are scaled up to match the
>> strongest data.
>>
>> Application of a elliptical resolution boundary is justified because
>> the resolution boundary from common integration programs (Denzo and
>> Mosflm for example) is spherical where diffraction for anisotropic
>> data is ellipsoidal. A spherical boundary would result in the
>> inclusion of numerous poorly measured reflections in the higher
>> resolution shells which effectively makes these data more noisy.
>> Imposing an ellipsoidal resolution boundary is equivalent to
>> removing noise from the higher resolution bins and is simply the
>> anisotropic equivalent of the normal resolution limit truncation.
>>
>> However, I was confused by the second and third steps.  The second
>> step of application of anisotropic scale factors is appropriate if
>> the refinement program does not include anisotropic scaling in its
>> calculation of Fc, however modern refinement programs do this. Pavel
>> Afonine touched on this in his CCP4BB general posting in response to
>> my original posting where he noted that "anisotropic scale factor[s]
>> that [are] part of the total structure factor take care of this" (<https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0909&L=CCP4BB&T=0&F=&S=&P=8362
>>> ).
>>
>> For the third step, applying a negative isotropic B-factor to modify
>> the Fo is equivalent to sharpening the peaks in your maps and this
>> can be useful.  However, applying the correction to Fo will also
>> result in an inappropriate decrease in the average temperature
>> factor of the resulting model.  Since B-factors are used as a
>> measure of the coordinate error of an atom, modifying your Fo means
>> these low B factors will tend to confuse the users of that model
>> into thinking its quality is better than it really is. If a sharper
>> map makes identification of model errors easier, the map can be
>> sharpened when it is calculated, without affecting the parameters in
>> the PDB file.  The latest versions of Coot, for example, allows you
>> to sharpen any map that it calculates.
>>
>> I brought these points to the attention of the Anisotropy Server
>> director (Michael Sawaya), who is now working to provide an option
>> to omit steps 2 and 3 for users who do not what their structure
>> factors modified.
>>
>> My thanks to everyone who responded to my original question, and to
>> Dale Tronrud and Michael Sawaya in particular for valuable  
>> discussion.

Harry
-- 
Dr Harry Powell, MRC Laboratory of Molecular Biology, MRC Centre,  
Hills Road, Cambridge, CB2 0QH