Saturday, 29 October 2011

Optimisation of weights


From: ‪<kavya
Date: 14 October 2011 06:12

Dear users,

Can the optimization of the X-ray weighing factor
and B-factor (overall wt) as mentioned in the paper
Acta Cryst. (2007). D63, 1274–1281 by Dr.Ian Tickel,
be used for the refinement of the data sets beyond
the resolution range mentioned in the paper: 1.33 -
2.55 Ang?

Also the structures that were used to optimize these
parameters were already solved and refined, so when
we are solving a new structure to what extent does the
model has to be built before starting the optimization?

Thanking you
With Regards
M. Kavyashree



----------
From: Ian Tickle

Hi Kavya

The resolutions of the structures mentioned in the paper were only
examples, the Rfree/-LLfree minimisation method (which are actually
due to Axel Brunger & Gerard Bricogne respectively) does not depend on
resolution.

If the structures are already solved & refined, you don't need to do
any model building, it should be within the radius of convergence with
the new weights - it's only a small adjustment after all.

Cheers

-- Ian


----------
From: ‪<kavya‬
Date: 14 October 2011 11:19



Respected Sir,

For one of the structures that I did optimisation
had values - (resolution of the data - 2.35Ang)

Before optimization- (Bfactor weight=1.0, X-ray Weight - auto)
R factor  0.2362
R free    0.2924
-LLfree   7521.8
rmsBOND   0.0160
zBOND     0.660

After optimisation- (B-factor weight=0.2, X-ray Weight - 0.08)
R factor  0.2327
R free    0.2882
-LLfree   7495.7
rmsBOND   0.0111
zBOND     0.460

Also can you tell me what is the limit for B-factor weight hat can be varied.
Sorry I just re-read your last email and realised and didn't read it
properly the first time.  But what I said still stands: you can of
course try to optimise the weights at an early stage (before adding
waters say), there's no harm doing that, but there's also not much
point since you'll have to do it all again with the complete model,
since adding a lot of waters will undoubtedly change the optimal
weights.  So I just leave the weight optimisation until the model is
complete.  As long as the initial weights are "in the same ball park",
so that your RMSZ(bonds) is around 0.5 for typical resolutions (a bit
lower for low resolution, a bit higher for very high resolution) it
won't affect interpretation of maps etc.

Cheers

-- Ian

On Fri, Oct 14, 2011 at 9:37 AM, Ian Tickle  wrote:
> It must be the same complete model that you refined previously, I
> doubt that it will give the correct answer if you leave out the waters
> for example.
> 
> You say "there was quite a difference".  Could you be more specific:
> what were the values of the weights, R factors and RMSZ(bonds/angles)
> before and after weight optimisation?
> 
> Cheers
> 
> -- Ian


----------
From: Ian Tickle


Hi your X-ray weight of .08 seems very small, the optimal value is
normally in the range 1 to 4 (I usually set it initially at the
median, i.e. 2.5).  But which weight keyword did you use "WEIGHT
MATRIX .08" or "WEIGHT AUTO .08" (the latter is I think undocumented,
so I'm guessing the first)?  Anyway I would strongly advise the
latter: the difference is that the MATRIX weight is on a completely
arbitrary scale, whereas the AUTO weight is at least relative to the
theoretical value of 1 (even though the optimal value may not be 1 in
practice, at least your initial guess will be in the same ball park).
Note that what Refmac calls "automatic weighting" is not the same as
what X-PLOR, CNS & phenix call "automatic weighting" (at least that's
my understanding).  "WEIGHT AUTO" in Refmac is the same as "WEIGHT
AUTO 10", whereas auto-weighting in X-PLOR corresponds to "WEIGHT AUTO
1" in Refmac.  Not surprisingly these give quite different results!

The optimal B factor weight is also around 1, see the paper for typical values.

I'm still not clear precisely what you meant by ""there was quite a
difference".  I don't see that big a difference between the 2 runs,
just a slight tightening up of the geometry.  Are you saying you see
big differences in the refined co-ordinates?  That would be a cause
for concern.

Cheers

-- Ian




From: ‪<kavya‬
Date: 14 October 2011 12:01


Respected Sir,

Yes, the weight mentioned in the paper was
weight matrix, but the one i used was the
option under "Refinement parameters- weighing
term (when auto weighing is turned off)".
But If I really wasnt to change the weight matrix
where should I change (in the code?)?

No, I dint mean a big difference, not in the
coordinates, but the values of R-factors and
other terms. I thought it was quite different.
So you mean that it is not of much concern?


Thanking you
With Regards
M. Kavyashree


----------
From: Ian Tickle
Date: 14 October 2011 12:03



No, the weights referred to in the paper are definitely the ones given
as "WEIGHT AUTO x".  I can say that with confidence because I never
use "WEIGHT MATRIX x" for reasons I explained.  I think what I said is
that it's _related_ to the matrix weight, which it obviously is by
some constant but unknown factor.

You don't have to change any code (that has been done for you!), but
if you're using CCP4I (sorry I don't so I can't give you precise
instructions), you probably have to edit the script before submission
(there should be a button in the "Run" menu for that).

Cheers

-- Ian
I would say that it's not a big difference, just tightening up of the
geometry, as I said.

Cheers

-- Ian








From: ‪<kavya
Date: 14 October 2011 08:42


Respected Sir,

Thank you for your clarification. I had adopted this
method recently. My doubt was if we have to optimize
these two parameters during refinement, should we have
the whole model along with water and ligands or only
protein with few water positioning is enough. The reason
why I am asking because there was quite a difference
when I refined the same structure without the optimization
and with optimization of these two parameters.

Thanking you
With regards
M. Kavyashree



  
   From: Ian Tickle
   Sent by: CCP4 bulletin board
   Date: 10/14/2011 12:34PM
   Subject: Re: [ccp4bb] Optimisation of weights

   Hi Kavya

   The resolutions of the structures mentioned in the paper were only
   examples, the Rfree/-LLfree minimisation method (which are actually
   due to Axel Brunger & Gerard Bricogne respectively) does not depend on
   resolution.

   If the structures are already solved & refined, you don't need to do
   any model building, it should be within the radius of convergence with
   the new weights - it's only a small adjustment after all.

   Cheers

   -- Ian

   On Fri, Oct 14, 2011 at 6:12 AM,  <kavya@ssl.serc.iisc.ernet.in> wrote:
   > Dear users,
   >
   > Can the optimization of the X-ray weighing factor
   > and B-factor (overall wt) as mentioned in the paper
   > Acta Cryst. (2007). D63, 1274–1281 by Dr.Ian Tickel,
   > be used for the refinement of the data sets beyond
   > the resolution range mentioned in the paper: 1.33 -
   > 2.55 Ang?
   >
   > Also the structures that were used to optimize these
   > parameters were already solved and refined, so when
   > we are solving a new structure to what extent does the
   > model has to be built before starting the optimization?
   >
   > Thanking you
   > With Regards
   > M. Kavyashree
   >
   >
   > --
   > This message has been scanned for viruses and
   > dangerous content by MailScanner, and is
   > believed to be clean.
   >



----------
From: Ian Ticklek


It must be the same complete model that you refined previously, I
doubt that it will give the correct answer if you leave out the waters
for example.

You say "there was quite a difference".  Could you be more specific:
what were the values of the weights, R factors and RMSZ(bonds/angles)
before and after weight optimisation?

Cheers

-- Ian


----------
From: Ian Tickle


Sorry I just re-read your last email and realised and didn't read it
properly the first time.  But what I said still stands: you can of
course try to optimise the weights at an early stage (before adding
waters say), there's no harm doing that, but there's also not much
point since you'll have to do it all again with the complete model,
since adding a lot of waters will undoubtedly change the optimal
weights.  So I just leave the weight optimisation until the model is
complete.  As long as the initial weights are "in the same ball park",
so that your RMSZ(bonds) is around 0.5 for typical resolutions (a bit
lower for low resolution, a bit higher for very high resolution) it
won't affect interpretation of maps etc.

Cheers

-- Ian


From: ‪<kavya
Date: 14 October 2011 12:34
Subject: [ccp4bb]

Respected Sir,

I am sorry I wrote it wrongly, its resolution-
independent X-ray weight rather. I use ccp4i
so the input for the weight is what i mentioned
previously-
"Refinement parameters- weighing term (when
auto weighing is turned off)" in refmac.

Thanking you
With Regads
M. Kavyashree


No comments:

Post a Comment