[phenixbb] Calculate F/sigma(F)

[Randy Read]

Hi Mirek,

It may or may not be possible in Phenix (I haven't really looked), but my question is why would you prefer to have the average F/sigma(F) when you could compute the much more meaningful average I/sigma(I)?  There are big problems with the meaning of F/sigma(F), and it depends on exactly how the intensities were converted to amplitudes.  Apart from French&Wilson, there is a variety of methods used, particularly to obtain a value for sigma(F).  When I/sigma(I) is large, the calculation is reasonably well-defined, and a Gaussian in I is close to a Gaussian in the derived F, but as the I/sigma(I) drops (which is exactly when you start to be concerned about the mean values of I/sigma(I) or F/sigma(F)!), the Gaussian approximation for the distribution of F becomes worse, and there's a serious problem when I<0.  This is where you have to use either a relatively arbitrary F=0, or apply the French&Wilson algorithm.  But the F&W algorithm has serious problems itself: first, it depends on having a good prior distribution (Wilson distribution) for the intensities, but many of the implementations ignore the systematic effect of anisotropy and, as far as I know, only Phaser takes account of the systematic effect of translational NCS; second, as sigma(I) becomes large compared to the Wilson expected intensity, F and sigma(F) converge on the mean and standard deviation of the Wilson distribution for amplitudes.  In the limit of infinite intensity measurement error (i.e. I/sigma(I)=0), the smallest F/sigma(F) from the F&W algorithm is about 1.913 for acentric reflections and 1.324 for centric reflections (http://journals.iucr.org/d/issues/2016/03/00/dz5382/index.html).

So, if there isn't a method to compute mean F/sigma(F) as a function of resolution in the Phenix package, I would argue that we shouldn't provide one because people really ought to be using mean I/sigma(I) or other intensity-based measures, and providing an option that shouldn't be used means that someone will nonetheless use it!

If you really need to do the calculation (e.g. to satisfy a referee, but then I would try first explaining to the referee why this isn't that helpful), you could do it in Bart Hazes' sftools program distributed with CCP4.  If you compute a new column with F/SIGF and then calculate the CC between that column and itself, the output reports the mean value of the column in resolution shells.  For instance, if you have an MTZ file with F in column 1 and SIGF in column 2, this should work:

   sftools
   read my.mtz
   calc col 3 = col 1 col 2 /
   correl col 3 3

If you don't like the default number of shells, you can use something like "correl col 3 3 shells 50".

Best wishes,

Randy

> On 18 Jan 2018, at 19:07, Cygler, Miroslaw <miroslaw.cygler at usask.ca> wrote:
> 
> Hi,
> I would like to calculate distribution of F/sigma(F) as a function of resolution for reflections in an mtz file. Is it possible to do it Phenix?
> Thanks,
> 
> 
> Mirek
> 
> 
> 
> _______________________________________________
> phenixbb mailing list
> phenixbb at phenix-online.org
> http://phenix-online.org/mailman/listinfo/phenixbb
> Unsubscribe: phenixbb-leave at phenix-online.org

------
Randy J. Read
Department of Haematology, University of Cambridge
Cambridge Institute for Medical Research      Tel: + 44 1223 336500
Wellcome Trust/MRC Building                   Fax: + 44 1223 336827
Hills Road                                    E-mail: rjr27 at cam.ac.uk
Cambridge CB2 0XY, U.K.                       www-structmed.cimr.cam.ac.uk

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://phenix-online.org/pipermail/phenixbb/attachments/20180119/a7c534bb/attachment.htm>