Hi,
I believe there used to be a constraint stopping refined SigmaA values from going negative, so I'm not sure how this is happening.
The basic problem here is that the correlation between Fo and Fc must be very close to zero, so that the refined SigmaA values go to zero (or should, in the case of the ones that go negative). It turns out that the likelihood is very insensitive to the exact value of SigmaA when the true value is close to zero, so it can become unstable.
Something odd must be happening in the density modification, which tends to increase the correlation between Fo and Fc even if the phases aren't improving (which is one of the reasons that the use of SIGMAA for back-transformed density-modified structure factors isn't really theoretically justified).
If the algorithm had some kind of smoothness restraint (like the algorithms used in modern refinement programs to determine SigmaA values from cross-validation data), it would probably behave better, but I suspect something more basic is going wrong.
Regards,
Randy Read
On 4 Jun 2010, at 11:09, Ian Tickle wrote:
> On Fri, Jun 4, 2010 at 2:28 AM, <[log in to unmask]> wrote:
>
>> I was using ccp4-6.0.2. I was using the following scripts:
>
>> sigmaa HKLIN test.mtz HKLOUT test-sigmaa.mtz << eof
>> title tt
>> labin FP=FP SIGFP=SIGFP FC=FP_atm PHIC=PHIC
>> labout DELFWT=DELFWT FWT=FWT WCMB=WCMB
>> symmetry P1211
>> END
>> eof
>
> I get the same results, i.e. some FOMs (about 6%) have gone negative,
> and about 75% are exactly zero, whether I use the 6.0.2 version or the
> current one. This message in the output is pertinent:
>
> Correlation between E**2's is non-positive -- SIGMAA set to 0.05
>
> Also this table (final iteration of optimization of sigma-A) is telling:
>
> Iteration Number 10 on SIGMAA
> OLD NEW MEAN
> D Limits SIGMAA Shift SIGMAA W
> 39.89 -- 8.74 0.00000 0.00000 0.00000 0.00000
> 8.74 -- 6.25 -0.59308 -0.00638 -0.59946 -0.45201
> 6.25 -- 5.13 0.66855 0.00000 0.66855 0.54835
> 5.13 -- 4.45 -0.61159 -0.00813 -0.61972 -0.49740
> 4.45 -- 3.99 0.71063 0.00000 0.71063 0.58634
> 3.99 -- 3.64 0.55934 0.00000 0.55934 0.43656
> 3.64 -- 3.37 0.00000 0.00000 0.00000 0.00000
> 3.37 -- 3.16 0.42551 0.00000 0.42551 0.32230
> 3.16 -- 2.98 0.00000 0.00000 0.00000 0.00000
> 2.98 -- 2.83 0.00000 0.00000 0.00000 0.00000
> 2.83 -- 2.69 0.00000 0.00000 0.00000 0.00000
> 2.69 -- 2.58 0.00000 0.00000 0.00000 0.00000
> 2.58 -- 2.48 0.00000 0.00000 0.00000 0.00000
> 2.48 -- 2.39 0.00000 0.00000 0.00000 0.00000
> 2.39 -- 2.31 0.00000 0.00000 0.00000 0.00000
> 2.31 -- 2.24 0.00000 0.00000 0.00000 0.00000
> 2.24 -- 2.17 0.00000 0.00000 0.00000 0.00000
> 2.17 -- 2.11 0.00000 0.00000 0.00000 0.00000
> 2.11 -- 2.05 0.00000 0.00000 0.00000 0.00000
> 2.05 -- 2.00 0.00000 0.00000 0.00000 0.00000
>
> Refinement of SIGMAA has not converged -- using current values
> Overall MEAN W is 0.05260
>
> DATA for SIGMAA PLOT
>
> STHOL**2 LN(SIGMAA)
> 0.008000 -0.402640
> 0.014211 -0.341601
> 0.017326 -0.580991
> 0.023544 -0.854478
>
> ln(SigmaA) vs sthol**2 should give straight line with
> slope = -26.3189*(mean square coordinate error) and intercept =
> 0.5*ln(sigmap/sigman).
>
>> Is that possible that the negative FOM is due to the large discrepancy
>> between FP and FC? Thanks for your kindly help!
>
> Yes I think that's it: the Fo/Fc agreement is essentially random at
> high resolution, so the sigma-A plot data only goes up to ~ 6.5 Ang
> and is meaningless, i.e. this is a case of the method being
> inappropriate rather than a program bug. Perhaps others who have
> experience of using this program in this situation can comment, my
> experience is only cases where agreement is good.
>
> Cheers
>
> -- Ian
------
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: [log in to unmask]
Cambridge CB2 0XY, U.K. www-structmed.cimr.cam.ac.uk
|