Re: [ccp4bb] Refining a crystal structure with (very) high solvent content

[Eleanor Dodson]

Well - I didnt want to give Dale a swelled head!
But Gerard and Jacob are both right..
E


On 4 June 2017 at 17:51, Keller, Jacob <kell...@janelia.hhmi.org> wrote:

> I would resolve this disagreement by repeating that "common sense is not
> so common." When I have seen a great scientist, or anyone with wisdom for
> that matter, I have seen the ability to demonstrate how complicated
> questions can be unravelled in a dazzlingly simple way, such that it almost
> seems trivial or common-sensical. Examples would include Richard Feynmann,
> the US Supreme Court Justices, and others. I would propose that Eleanor
> meant this kind of uncommon common sense.
>
> JPK
>
> -Original Message-
> From: CCP4 bulletin board [mailto:CCP4BB@JISCMAIL.AC.UK] On Behalf Of
> Gerard Bricogne
> Sent: Sunday, June 04, 2017 12:07 PM
> To: CCP4BB@JISCMAIL.AC.UK
> Subject: Re: [ccp4bb] Refining a crystal structure with (very) high
> solvent content
>
> Dear Eleanor,
>
>  I think this is too faint a praise for Dale. What he shows in his
> reply is not just common sense, but knowledge and understanding of the
> fundamentals. You can't do good science with common sense alone, and in our
> field common sense will not be of much help if you do not understand the
> Fourier transform well enough, for example.
>
>  I would venture to guess that 95+% of crystallographers are in the
> unquestioned habit of making the same conceptual error that Dale has
> pointed out, viz. mistaking the rmsd of the map (which is a unit of
> contrast) for the standard deviation of a noise level in the map.
> The latter quantity has nothing to do with the former, as has been pointed
> out many times.
>
>  The problem is that this confusion is enshrined in the default values
> of certain parameters in display programs and scripts, that are assumed
> (not by their authors, but by almost everybody else) to embody all the
> common sense we need :-) .
>
>
>  With best wishes,
>
>   Gerard.
>
> --
> On Sun, Jun 04, 2017 at 03:36:29PM +0100, Eleanor Dodson wrote:
> > Thank you Dale! You talk so much common sense..
> > Eleanor
> >
> > On 2 June 2017 at 23:30, Dale Tronrud <de...@daletronrud.com> wrote:
> >
> > > On 6/2/2017 1:42 PM, wtempel wrote:
> > > > Hello all,
> > > > crystals with high solvent content tend to diffract poorly, at
> > > > least according to intuition. Several years ago we solved a
> > > > structure
> > > > <http://www.rcsb.org/pdb/explore/explore.do?structureId=2h58> that
> > > > appeared to buck that trend with a solvent content of ≈0.8 and
> > > > resolution beyond 2 Å, per merging statistics and visibility of
> spots on diffraction images.
> > > > I would welcome my colleagues’ opinions as to why I might observe
> > > > the
> > > > following:
> > > >
> > > >  1. Paired refinement (similar to Fig. 1 in Karplus
> > > > <http://doi.org10.1126/science.1218231>) indicates that adding
> any
> > > > higher resolution data beyond 3.4 Å, the lowest high resolution
> > > > cut-off limit I tried, does not improve R-factors at the common
> > > > lower resolution cutoff. Yes, diffraction is anisotropic in this
> > > > case, but seemingly not to that extent. I hesitate to “throw out”
> > > > all data beyond 3.4 Å, or whatever lower resolution cut-off I
> > > > might
> > > try.
> > > >  2. The Fo-Fc map, when countoured at ± 3 rmsd, includes many more
> > > > (uninterpretable) features than I would expect after refinement
> to
> > > > residuals in the mid-to-lower twenties. For expected map
> appearance,
> > > > I had to crank up the coutour level to > 5 rmsd, like in the
> > > > attached screenshot of the ADP·Mg^++ omit map.
> > >
> > >This is one of the prime examples of the failure of describing
> > > contour levels in terms of "sigma".  First, the number you are using
> > > is not a "standard deviation" or any other measure of the error
> > > level of the map but is simply the rms value of the map.  If you
> > > calculate the rms of a difference map where 80% of the unit cell is
> > > bulk solvent, and therefore flat, you will, of course, get a much
> > > smaller number than if the unit cell contained 80% protein with all
> > > the the expected difference map features that come from a model with
> > > an R value of ~20%.  Then when you contour at three time

Re: [ccp4bb] Refining a crystal structure with (very) high solvent content

[Keller, Jacob]

I would resolve this disagreement by repeating that "common sense is not so 
common." When I have seen a great scientist, or anyone with wisdom for that 
matter, I have seen the ability to demonstrate how complicated questions can be 
unravelled in a dazzlingly simple way, such that it almost seems trivial or 
common-sensical. Examples would include Richard Feynmann, the US Supreme Court 
Justices, and others. I would propose that Eleanor meant this kind of uncommon 
common sense.

JPK

-Original Message-
From: CCP4 bulletin board [mailto:CCP4BB@JISCMAIL.AC.UK] On Behalf Of Gerard 
Bricogne
Sent: Sunday, June 04, 2017 12:07 PM
To: CCP4BB@JISCMAIL.AC.UK
Subject: Re: [ccp4bb] Refining a crystal structure with (very) high solvent 
content

Dear Eleanor,

 I think this is too faint a praise for Dale. What he shows in his reply is 
not just common sense, but knowledge and understanding of the fundamentals. You 
can't do good science with common sense alone, and in our field common sense 
will not be of much help if you do not understand the Fourier transform well 
enough, for example.

 I would venture to guess that 95+% of crystallographers are in the 
unquestioned habit of making the same conceptual error that Dale has pointed 
out, viz. mistaking the rmsd of the map (which is a unit of contrast) for the 
standard deviation of a noise level in the map.
The latter quantity has nothing to do with the former, as has been pointed out 
many times.

 The problem is that this confusion is enshrined in the default values of 
certain parameters in display programs and scripts, that are assumed (not by 
their authors, but by almost everybody else) to embody all the common sense we 
need :-) .


 With best wishes,
 
  Gerard.

--
On Sun, Jun 04, 2017 at 03:36:29PM +0100, Eleanor Dodson wrote:
> Thank you Dale! You talk so much common sense..
> Eleanor
> 
> On 2 June 2017 at 23:30, Dale Tronrud <de...@daletronrud.com> wrote:
> 
> > On 6/2/2017 1:42 PM, wtempel wrote:
> > > Hello all,
> > > crystals with high solvent content tend to diffract poorly, at 
> > > least according to intuition. Several years ago we solved a 
> > > structure 
> > > <http://www.rcsb.org/pdb/explore/explore.do?structureId=2h58> that 
> > > appeared to buck that trend with a solvent content of ≈0.8 and 
> > > resolution beyond 2 Å, per merging statistics and visibility of spots on 
> > > diffraction images.
> > > I would welcome my colleagues’ opinions as to why I might observe 
> > > the
> > > following:
> > >
> > >  1. Paired refinement (similar to Fig. 1 in Karplus
> > > <http://doi.org10.1126/science.1218231>) indicates that adding any
> > > higher resolution data beyond 3.4 Å, the lowest high resolution
> > > cut-off limit I tried, does not improve R-factors at the common
> > > lower resolution cutoff. Yes, diffraction is anisotropic in this
> > > case, but seemingly not to that extent. I hesitate to “throw out”
> > > all data beyond 3.4 Å, or whatever lower resolution cut-off I 
> > > might
> > try.
> > >  2. The Fo-Fc map, when countoured at ± 3 rmsd, includes many more
> > > (uninterpretable) features than I would expect after refinement to
> > > residuals in the mid-to-lower twenties. For expected map appearance,
> > > I had to crank up the coutour level to > 5 rmsd, like in the
> > > attached screenshot of the ADP·Mg^++ omit map.
> >
> >This is one of the prime examples of the failure of describing 
> > contour levels in terms of "sigma".  First, the number you are using 
> > is not a "standard deviation" or any other measure of the error 
> > level of the map but is simply the rms value of the map.  If you 
> > calculate the rms of a difference map where 80% of the unit cell is 
> > bulk solvent, and therefore flat, you will, of course, get a much 
> > smaller number than if the unit cell contained 80% protein with all 
> > the the expected difference map features that come from a model with 
> > an R value of ~20%.  Then when you contour at three times this 
> > absurdly small number you will see all sorts of features you are not 
> > used to seeing.  Selecting a contour level based on the e/A^3 is 
> > much less sensitive to the amount of solvent in the crystal is gives much 
> > more consistent results.
> >
> > Dale Tronrud
> > >
> > > Could these observations be linked to the high solvent content? 
> > > (1) A high solvent content structure has a higher-than-average 
> > > observation-to-parameter ratio, suffici

Re: [ccp4bb] Refining a crystal structure with (very) high solvent content

[Gerard Bricogne]

Dear Eleanor,

 I think this is too faint a praise for Dale. What he shows in his
reply is not just common sense, but knowledge and understanding of the
fundamentals. You can't do good science with common sense alone, and
in our field common sense will not be of much help if you do not
understand the Fourier transform well enough, for example.

 I would venture to guess that 95+% of crystallographers are in
the unquestioned habit of making the same conceptual error that Dale
has pointed out, viz. mistaking the rmsd of the map (which is a unit
of contrast) for the standard deviation of a noise level in the map.
The latter quantity has nothing to do with the former, as has been
pointed out many times.

 The problem is that this confusion is enshrined in the default
values of certain parameters in display programs and scripts, that are
assumed (not by their authors, but by almost everybody else) to embody
all the common sense we need :-) .


 With best wishes,
 
  Gerard.

--
On Sun, Jun 04, 2017 at 03:36:29PM +0100, Eleanor Dodson wrote:
> Thank you Dale! You talk so much common sense..
> Eleanor
> 
> On 2 June 2017 at 23:30, Dale Tronrud  wrote:
> 
> > On 6/2/2017 1:42 PM, wtempel wrote:
> > > Hello all,
> > > crystals with high solvent content tend to diffract poorly, at least
> > > according to intuition. Several years ago we solved a structure
> > >  that
> > > appeared to buck that trend with a solvent content of ≈0.8 and
> > > resolution beyond 2 Å, per merging statistics and visibility of spots on
> > > diffraction images.
> > > I would welcome my colleagues’ opinions as to why I might observe the
> > > following:
> > >
> > >  1. Paired refinement (similar to Fig. 1 in Karplus
> > > ) indicates that adding any
> > > higher resolution data beyond 3.4 Å, the lowest high resolution
> > > cut-off limit I tried, does not improve R-factors at the common
> > > lower resolution cutoff. Yes, diffraction is anisotropic in this
> > > case, but seemingly not to that extent. I hesitate to “throw out”
> > > all data beyond 3.4 Å, or whatever lower resolution cut-off I might
> > try.
> > >  2. The Fo-Fc map, when countoured at ± 3 rmsd, includes many more
> > > (uninterpretable) features than I would expect after refinement to
> > > residuals in the mid-to-lower twenties. For expected map appearance,
> > > I had to crank up the coutour level to > 5 rmsd, like in the
> > > attached screenshot of the ADP·Mg^++ omit map.
> >
> >This is one of the prime examples of the failure of describing
> > contour levels in terms of "sigma".  First, the number you are using is
> > not a "standard deviation" or any other measure of the error level of
> > the map but is simply the rms value of the map.  If you calculate the
> > rms of a difference map where 80% of the unit cell is bulk solvent, and
> > therefore flat, you will, of course, get a much smaller number than if
> > the unit cell contained 80% protein with all the the expected difference
> > map features that come from a model with an R value of ~20%.  Then when
> > you contour at three times this absurdly small number you will see all
> > sorts of features you are not used to seeing.  Selecting a contour level
> > based on the e/A^3 is much less sensitive to the amount of solvent in
> > the crystal is gives much more consistent results.
> >
> > Dale Tronrud
> > >
> > > Could these observations be linked to the high solvent content? (1) A
> > > high solvent content structure has a higher-than-average
> > > observation-to-parameter ratio, sufficiently high even when limited to
> > > stronger, low-resolution reflections? (2) Map normalization may not be
> > > attuned to such high solvent content?
> > > I am interested in analyzing the automated decision-making of the
> > > PDB-REDO of this entry , such as
> > > paired refinement results and selection of ADP model. Should I find this
> > > information in the “All files (compressed)” archive
> > > ? The “fully
> > > optimized structure’
> > >  shows |ANISOU|
> > > cards and |NUMBER OF TLS GROUPS : NULL|. Does this mean that individual
> > > ADPs have been refined anisotropically?
> > > Looking forward to your insights,
> > > Wolfram Tempel
> > >
> > > ​

Re: [ccp4bb] Refining a crystal structure with (very) high solvent content

[Eleanor Dodson]

Thank you Dale! You talk so much common sense..
Eleanor

On 2 June 2017 at 23:30, Dale Tronrud  wrote:

> On 6/2/2017 1:42 PM, wtempel wrote:
> > Hello all,
> > crystals with high solvent content tend to diffract poorly, at least
> > according to intuition. Several years ago we solved a structure
> >  that
> > appeared to buck that trend with a solvent content of ≈0.8 and
> > resolution beyond 2 Å, per merging statistics and visibility of spots on
> > diffraction images.
> > I would welcome my colleagues’ opinions as to why I might observe the
> > following:
> >
> >  1. Paired refinement (similar to Fig. 1 in Karplus
> > ) indicates that adding any
> > higher resolution data beyond 3.4 Å, the lowest high resolution
> > cut-off limit I tried, does not improve R-factors at the common
> > lower resolution cutoff. Yes, diffraction is anisotropic in this
> > case, but seemingly not to that extent. I hesitate to “throw out”
> > all data beyond 3.4 Å, or whatever lower resolution cut-off I might
> try.
> >  2. The Fo-Fc map, when countoured at ± 3 rmsd, includes many more
> > (uninterpretable) features than I would expect after refinement to
> > residuals in the mid-to-lower twenties. For expected map appearance,
> > I had to crank up the coutour level to > 5 rmsd, like in the
> > attached screenshot of the ADP·Mg^++ omit map.
>
>This is one of the prime examples of the failure of describing
> contour levels in terms of "sigma".  First, the number you are using is
> not a "standard deviation" or any other measure of the error level of
> the map but is simply the rms value of the map.  If you calculate the
> rms of a difference map where 80% of the unit cell is bulk solvent, and
> therefore flat, you will, of course, get a much smaller number than if
> the unit cell contained 80% protein with all the the expected difference
> map features that come from a model with an R value of ~20%.  Then when
> you contour at three times this absurdly small number you will see all
> sorts of features you are not used to seeing.  Selecting a contour level
> based on the e/A^3 is much less sensitive to the amount of solvent in
> the crystal is gives much more consistent results.
>
> Dale Tronrud
> >
> > Could these observations be linked to the high solvent content? (1) A
> > high solvent content structure has a higher-than-average
> > observation-to-parameter ratio, sufficiently high even when limited to
> > stronger, low-resolution reflections? (2) Map normalization may not be
> > attuned to such high solvent content?
> > I am interested in analyzing the automated decision-making of the
> > PDB-REDO of this entry , such as
> > paired refinement results and selection of ADP model. Should I find this
> > information in the “All files (compressed)” archive
> > ? The “fully
> > optimized structure’
> >  shows |ANISOU|
> > cards and |NUMBER OF TLS GROUPS : NULL|. Does this mean that individual
> > ADPs have been refined anisotropically?
> > Looking forward to your insights,
> > Wolfram Tempel
> >
> > ​
>

Re: [ccp4bb] Refining a crystal structure with (very) high solvent content

[Dale Tronrud]

On 6/2/2017 1:42 PM, wtempel wrote:
> Hello all,
> crystals with high solvent content tend to diffract poorly, at least
> according to intuition. Several years ago we solved a structure
>  that
> appeared to buck that trend with a solvent content of ≈0.8 and
> resolution beyond 2 Å, per merging statistics and visibility of spots on
> diffraction images.
> I would welcome my colleagues’ opinions as to why I might observe the
> following:
> 
>  1. Paired refinement (similar to Fig. 1 in Karplus
> ) indicates that adding any
> higher resolution data beyond 3.4 Å, the lowest high resolution
> cut-off limit I tried, does not improve R-factors at the common
> lower resolution cutoff. Yes, diffraction is anisotropic in this
> case, but seemingly not to that extent. I hesitate to “throw out”
> all data beyond 3.4 Å, or whatever lower resolution cut-off I might try.
>  2. The Fo-Fc map, when countoured at ± 3 rmsd, includes many more
> (uninterpretable) features than I would expect after refinement to
> residuals in the mid-to-lower twenties. For expected map appearance,
> I had to crank up the coutour level to > 5 rmsd, like in the
> attached screenshot of the ADP·Mg^++ omit map.

   This is one of the prime examples of the failure of describing
contour levels in terms of "sigma".  First, the number you are using is
not a "standard deviation" or any other measure of the error level of
the map but is simply the rms value of the map.  If you calculate the
rms of a difference map where 80% of the unit cell is bulk solvent, and
therefore flat, you will, of course, get a much smaller number than if
the unit cell contained 80% protein with all the the expected difference
map features that come from a model with an R value of ~20%.  Then when
you contour at three times this absurdly small number you will see all
sorts of features you are not used to seeing.  Selecting a contour level
based on the e/A^3 is much less sensitive to the amount of solvent in
the crystal is gives much more consistent results.

Dale Tronrud
> 
> Could these observations be linked to the high solvent content? (1) A
> high solvent content structure has a higher-than-average
> observation-to-parameter ratio, sufficiently high even when limited to
> stronger, low-resolution reflections? (2) Map normalization may not be
> attuned to such high solvent content?
> I am interested in analyzing the automated decision-making of the
> PDB-REDO of this entry , such as
> paired refinement results and selection of ADP model. Should I find this
> information in the “All files (compressed)” archive
> ? The “fully
> optimized structure’
>  shows |ANISOU|
> cards and |NUMBER OF TLS GROUPS : NULL|. Does this mean that individual
> ADPs have been refined anisotropically?
> Looking forward to your insights,
> Wolfram Tempel
> 
> ​

Re: [ccp4bb] Refining a crystal structure with (very) high solvent content

[Robbie Joosten]

Hi Wolfram,

We recently moved pdb-redo to a new domain, so the entry you are interested in 
is at https://pdb-redo.eu/db/2h58
This entry at 1.85A resolution was indeed refined with anisotropic B-factors 
based on the Hamilton R ratio test. There was no paired refinement for this 
entry. We only do that if the deposited data extends to a resolution higher 
than what was used in the refinement of the PDB entry.
Of course you can use our server to do paired refinement for your dataset. It 
works best when your model is already in a good state.

Cheers,
Robbie

Sent from my Windows 10 phone

Van: wtempel<mailto:wtem...@gmail.com>
Verzonden: vrijdag 2 juni 2017 22:44
Aan: CCP4BB@JISCMAIL.AC.UK<mailto:CCP4BB@JISCMAIL.AC.UK>
Onderwerp: [ccp4bb] Refining a crystal structure with (very) high solvent 
content


Hello all,
crystals with high solvent content tend to diffract poorly, at least according 
to intuition. Several years ago we solved a 
structure<http://www.rcsb.org/pdb/explore/explore.do?structureId=2h58> that 
appeared to buck that trend with a solvent content of ≈0.8 and resolution 
beyond 2 Å, per merging statistics and visibility of spots on diffraction 
images.
I would welcome my colleagues’ opinions as to why I might observe the following:

  1.  Paired refinement (similar to Fig. 1 in 
Karplus<http://doi.org10.1126/science.1218231>) indicates that 
adding any higher resolution data beyond 3.4 Å, the lowest high resolution 
cut-off limit I tried, does not improve R-factors at the common lower 
resolution cutoff. Yes, diffraction is anisotropic in this case, but seemingly 
not to that extent. I hesitate to “throw out” all data beyond 3.4 Å, or 
whatever lower resolution cut-off I might try.
  2.  The Fo-Fc map, when countoured at ± 3 rmsd, includes many more 
(uninterpretable) features than I would expect after refinement to residuals in 
the mid-to-lower twenties. For expected map appearance, I had to crank up the 
coutour level to > 5 rmsd, like in the attached screenshot of the ADP·Mg++ omit 
map.

Could these observations be linked to the high solvent content? (1) A high 
solvent content structure has a higher-than-average observation-to-parameter 
ratio, sufficiently high even when limited to stronger, low-resolution 
reflections? (2) Map normalization may not be attuned to such high solvent 
content?
I am interested in analyzing the automated decision-making of the PDB-REDO of 
this entry<http://www.cmbi.ru.nl/pdb_redo/h5/2h58>, such as paired refinement 
results and selection of ADP model. Should I find this information in the “All 
files (compressed)” 
archive<http://www.cmbi.ru.nl/pdb_redo/cgi-bin/zipper.pl?id=2h58>? The “fully 
optimized structure’<http://www.cmbi.ru.nl/pdb_redo/h5/2h58/2h58_final.pdb> 
shows ANISOU cards and NUMBER OF TLS GROUPS : NULL. Does this mean that 
individual ADPs have been refined anisotropically?
Looking forward to your insights,
Wolfram Tempel

​