Dear Eleanor and Kay,
Thanks for pointing out that the difference maps does not show model
bias due to "fill-in".
The difference maps generated with "fill-in" and without (mapcalculate
free include) look somewhat different for another dataset I tried. But,
these differences are small.
Notably, 2fofc map caluculated with "mapcalculate free include" look
weird for this dataset. This outcome is somewhat expected due to low
completeness in lower shells (yes, this dataset is problematic). In the
first dataset, which has good low res completeness, omit maps looked
somewhat better, as I wrote before.
My conclusion is that "no fill-in" option might be tried at some
stages, but with caution, especially for datasets with poor low res
completeness.
Many thanks to everyone!
Ivan
With best regards,
Ivan Shabalin, Ph.D.
Research Scientist,
Department of Molecular Physiology and Biological Physics,
University of Virginia,
1340 Jefferson Park Avenue, Pinn Hall,Room 4223,
Charlottesville, VA 22908
https://www.linkedin.com/in/shabalinig/
https://minorlab.org/person/ivan_s/
On 8/6/19 09:39, Eleanor Dodson wrote:
There are certainly some problems with REFMAC after an anisotropy
correction. The FreeR flags and D corrections are organised in spherical
shells, whereas the D aniso volume is elliptical..
As Kay points out the REFMAC map coefficient for a difference map is 0
if there is no measurement so there is no bias in those maps.
the 2mFo-DFc maps do set the map coefficient to D Fc if there is no
measurement, potentially introducing bias.
The thinking is that when a large low resolution term is missing it is
better to approximate it to something realistic, rather than set it to 0.
For high weak resolution data D Fc will also probably be insignificant
and there would be very little gain by including it, and the possibility
of introducingbias..
However nowadays the data collection techniques usually provide pretty
complete data at low resolution,
On Tue, 6 Aug 2019 at 07:44, Kay Diederichs
<[email protected] <mailto:[email protected]>>
wrote:
Dear Ivan,
one thing I forgot to mention: I think the difference maps should
not show model bias due to "fill-in". This is because the term
D*Fcalc that is filled in as a replacement for m*Fobs is just
compensated by the term D*Fc that is subtracted when forming the
m*Fobs - D*Fcalc difference coefficients.
This means that when building the model, you can let yourself be
guided primarily by the difference maps. These will not suffer from
model bias due to fill-in.
I like the idea of the "shaping the MTZ file" that Robbie suggests,
but I still need to work out the proper sftools commands, like those
at http://staraniso.globalphasing.org/test_set_flags_about.html .
best,
Kay
On Mon, 5 Aug 2019 17:44:56 -0400, Ivan Shabalin
<[email protected]
<mailto:[email protected]>> wrote:
>Dear Kay,
>
>Thanks a lot for your answers!
>
>To my best understanding, REFMAC does not have an option of for
>restoring reflections only in certain resolution shells. But, it
should
>not be a problem for datasets with good completeness in low
resolution
>shells. Also, refinement against intensities is available only for
twin
>refinement.
>
>I will take this as a conclusion for datasets with good low
resolution
>completeness: "even if the visual effect of weak reflections on
the map
>may be low, the errors in the model coordinates will be less if
the weak
>amplitudes are used in refinement"
>
>Thanks!
>
>Ivan
>
>
>With best regards,
>Ivan Shabalin, Ph.D.
>Research Scientist,
>Department of Molecular Physiology and Biological Physics,
>University of Virginia,
>1340 Jefferson Park Avenue, Pinn Hall,Room 4223,
>Charlottesville, VA 22908
>https://www.linkedin.com/in/shabalinig/
>https://minorlab.org/person/ivan_s/
>
>On 8/3/19 04:32, Kay Diederichs wrote:
>> Dear Ivan,
>>
>> On Thu, 1 Aug 2019 22:10:36 -0400, Ivan Shabalin
<[email protected]
<mailto:[email protected]>> wrote:
>>
>>> Dear CCP4BB,
>>>
>>> There seems to be a general consensus for extending data to higher
>>> resolution to include as much meaningful data as possible.
"Meaningful"
>>> can be defined in different ways. I heard/read opinions such as 0.5
>>> CC1/2, 0.3 CC1/2, 0.15 CC1/2,
>>
>> This is numerology - why not 0.33333 or 0.12345? The EM
community has agreed on the "gold standard" of 0.143 for FSC which
has a similar definition as CC1/2 - this value is chosen because the
quantity analogous to CC* is then 0.5 !
>>
>>> and stepped (paired) refinement. The
>>> latter seems to be one of the most rigorous options according
to many
>>> crystallographers.
>>>
>>> Including more data sounds like a good thing, but, it sometimes
results
>>> in low completeness in high resolution shells. As far as i
understand,
>>> this may result from:
>>>
>>> a) anisotropic diffraction (if a software cuts of resolution in
>>> non-isotropic way)
>>>
>>> b) sub-optimal data collection (e.g. due to limitations of the
>>> instrument, such as minimum detector distance allowed, absence
of kappa,
>>> limits on oscillation range)
>>>
>>> In the commonly referred paper, the completeness is 96% in the
highest
>>> shell (Karplus, P. A., & Diederichs, K. (2012). Linking
crystallographic
>>> model and data quality. Science (New York, N.Y.), 336(6084),
1030–1033.)
>>> In other words, these tests were performed for an almost
complete dataset.
>>>
>>> I used to think that more data is always better, but, as I learned
>>> recently from Clemens Vonrhein, the resulting low completeness
may cause
>>> model bias in the maps.
>>
>> ... due to "fill-in" of missing reflections which is performed
by Refmac and phenix.refine
(https://www.phenix-online.org/documentation/faqs/refine.html#general);
don't know about other programs.
>>
>>> Indeed, REFMAC by default tries to restore missing reflections,
which
>>> are approximated as DFc
>>>
(https://www2.mrc-lmb.cam.ac.uk/groups/murshudov/content/refmac/refmac_keywords.html).
>>>
>>>
>>> We tried using the keyword "mapcalculate free include" and
"mapcalculate
>>> free exclude" for one of our structures (~1.3A, P1), and it did
seem to
>>> improve the maps a little - we saw more meaningful features.
>>>
>>> But, I still have several questions:
>>>
>>> 1) Does using "mapcalculate free include" in REFMAC represent a
sound
>>> solution to this problem? Does this "no fill-in at all" solution
>>> constitute a significant problem?
>>
>> I think it would be desirable to "fill in" only the reflections
missing at low resolution (e.g. in your case to 3A); that would
improve the maps but decrease the model bias. Not sure if any
program supports this.
>>
>>>
>>> 2) Are there any other concerns about using data with low
completeness
>>> in highest shells?
>>
>> no.
>>
>> To the contrary, even if the visual effect of weak reflections
on the map may be low, the errors in the model coordinates will be
less if the weak amplitudes (or even better, intensities - once the
refinement programs use these) are used as restraints in refinement.
>>
>>>
>>> 3) STARANISO website suggests a way of handling this problem
>>> (http://staraniso.globalphasing.org/test_set_flags_about.html).
But,
>>> would not REFMAC "fill-in" all the reflections for map coefficients
>>> calculation to isotropic completeness anyway?
>>>
>>> 4) What is your personal approach to handling this issue? Is there
>>> completeness value in the last shell that is too low to include
it in
>>> Table 1?
>>
>> To me, the rule "the completeness in high-resolution shell is
too low if it is less than XX %" appears not to be based on any
appropriate concept. If there is good information in that resolution
shell, it should be used no matter what the completeness is.
Practically, this means that one has to devote a lot of attention to
doing the experiment right, by choosing an appropriate
crystal-detector distance. As always, attempts to computationally
rectify problems that result from a poor experiment have their own
caveats.
>>
>> Thanks for bringing this up!
>>
>> Kay
>>
>>>
>>> Many thanks,
>>>
>>> Ivan
>>>
>>> With best regards,
>>> Ivan Shabalin, Ph.D.
>>> Research Scientist,
>>> Department of Molecular Physiology and Biological Physics,
>>> University of Virginia,
>>> 1340 Jefferson Park Avenue, Pinn Hall,Room 4223,
>>> Charlottesville, VA 22908
>>> https://www.linkedin.com/in/shabalinig/
>>> https://minorlab.org/person/ivan_s/
>>>
>>>
>>>
>>>
>>>
>>> ----
>>>
>>>
########################################################################
>>>
>>> To unsubscribe from the CCP4BB list, click the following link:
>>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCP4BB&A=1
>>
>>
########################################################################
>>
>> To unsubscribe from the CCP4BB list, click the following link:
>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCP4BB&A=1
>>
>
>----
>
>########################################################################
>
>To unsubscribe from the CCP4BB list, click the following link:
>https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCP4BB&A=1
########################################################################
To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCP4BB&A=1
------------------------------------------------------------------------
To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCP4BB&A=1
----
########################################################################
To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCP4BB&A=1
