Frank,

This has features in common with my proposal at the 2011 CCP4SW to use
<rho_obs / SU(rho_diff)> as a model reliability (or precision) metric
(now implemented in EDSTATS).  This should not be confused with my
proposed accuracy metric which uses a chi-square test based on the
values of rho_diff / SU(rho_diff).

The main difference between the precision metric proposed in this
paper (i.e. presumably rho_obs / SU(rho_obs)) and mine is that I use
SU(rho_diff) in place of SU(rho_obs) because it's much easier to get
the former from a QQ plot of rho_diff (making the not unreasonable
assumption that if only random errors are present then rho_diff has a
normal error distribution).  I also tried to compute SU(rho_obs)
directly but the results I got didn't look right so I didn't pursue
it.

Of course dividing rho_obs by the SU is the easy bit: determining the
SU values in the first place is the hard part (as I said, I've
tried!), and I guess that is what this paper is all about.  If they
have a good way of getting SU(rho_obs) by FFT then that's great - I
can use it in EDSTATS!

Ideally I think COOT should implement the whole package of e.d.
metrics (RSR, RSCC as well as the ones described above) so people can
decide for themselves what works best.

I haven't actually been able to read their paper yet so I can't answer
these questions.  I have to go through our individual reprint ordering
service which takes 4-5 days to get it, so hopefully I'll have more
info next week.

Cheers

-- Ian

On 14 June 2012 19:34, Frank von Delft <[email protected]> wrote:
> Proposition:  what the paper below describes (map variance) would be rather
> useful to access routinely (e.g. in coot etc), given that it describes how
> to calculate by FFT.
>
> Discuss.  (No, I've not read any of it carefully.)
>
>
>
>
> Acta Cryst. (2012). A68, 464-469  [ doi:10.1107/S0108767312015772 ]
>
> A practical study of the electron-density-map variance
>
> C. Giacovazzo and A. Mazzone
>
> An algorithm is described for the calculation of the variance of the
> electron density via fast Fourier transform. The main features of the ratio
> electron density/standard deviation are described; it is shown that such a
> ratio may be considered a useful criterion for estimating, on an absolute
> scale, the quality of an electron-density map, no matter the quality of the
> model and the data resolution. In some way, it is a good estimate of the
> confidence one should have in the reliability of the proposed structure. The
> ratio was tested both for observed and for difference Fourier syntheses, in
> order to control the usefulness of the criterion for the most popular maps
> in crystallography.

Reply via email to