[phenixbb] model vs Wilson b-factor

[Dale Tronrud]

   It's also true that the Wilson B calculation assumes that all the B factors
in the crystal are the same - which is also far from the true in most macromolecular
crystals.  A person who holds to the practice of aggressively building water
molecules and loops will create a model with a higher average B than one who
uses the same data but is more restrained.  The Wilson B will, of course, be
unchanged.

   If you have a protein that is equally ordered throughout and you do not
build in weak water molecules and the crystal diffracts high enough to allow
a reasonably accurate calculation of the Wilson B, your average B and and the
Wilson B should be close to each other.  If your protein has mobile loops,
which most lower resolution crystals do (and most higher resolution crystals
for that matter) then your average B will be larger than your Wilson B.

   Since the Wilson B and the average B are such different quantities I don't
believe there are any useful conclusions that can be made by comparing the
two.  If you want to see if your model is consistent with your Wilson B you
should calculate F squared values from it and calculate a Wilson B from those.
If the calculated Wilson B doesn't match the observed Wilson B your model has
a serious problem, but I expect that every refinement program will produce
models that match quite closely - even models that are quite wrong will at
least match the Wilson B.  A discrepancy between calculated and observed
Wilson B causes a horrible increase in R values (both kinds) which is easily
fixed by the refinement program adjusting whatever B values are being refined.

Dale Tronrud

On 04/21/10 15:01, MARTYN SYMMONS wrote:
> Perhaps worth pointing out that that Wilson B is the based on the assumption of randomly distributed atoms. This is not at all how proteins are, and in particular secondary structures give a preponderance of spacings in the 4 angstrom-ish region and a peak of mean intensity in these shells. For this reason the apparent fall-off in the Wilson in this resolution range is steeper as you are falling down off of the peak due to this seondary structure giving favoured spacings that produce a deviation from randomness in this resolution range. So it will be dependent on the secondary structure of the an individual protein . So the Wilson gets about right when you deal with spacings that tend to be unbiased by secondary structure - which unfortunately is the bit that is missing in the low resolution crystal data. Wilson fall off in low resolution looks steep because the random assumption is invalid. 
> 
> Maybe you can guess the secondary content of your protein from where the bump is in the Wilson plot - beta gives a bulge in the 4 ang region - alpha in the 5 to 9 ang region. 
> 
> all the best
>                Martyn 
> 
> Martyn Symmons
> Cambridge
> 
> 
> 
> 
> ----- Original Message ----
> From: Pavel Afonine <PAfonine at lbl.gov>
> To: Gino.Cingolani at jefferson.edu; PHENIX user mailing list <phenixbb at phenix-online.org>
> Sent: Wednesday, 21 April, 2010 18:56:47
> Subject: Re: [phenixbb] model vs Wilson b-factor
> 
> Hi Gino,
> 
> here are a few points:
> 
> - my understanding (please correct me if I'm wrong) is that the accuracy of Wilson B estimate drops with the resolution: lower the resolution, less accurate is the estimate;
> 
> - Wilson B is not a given calculated value - it's just an estimate;
> 
> - the total atomic B-factor includes the trace of overall anisotropic scale matrix (see Fmodel formula for the total model structure factor: Fmodel = scale_overall * exp(-h*U_overall*ht) * (Fcalc + k_sol * exp(-B_sol*s^2) * Fmask) ). You can try to disable this and see if this was the cause (use "apply_back_trace_of_b_cart=true" keyword for this).
> 
> - the things you "tried to resolve this discrepancy" will unlikely to change the average B-factor;
> 
> - assuming that you used the proper model parameterization and refinement strategy given your model and data quality,  I would just accept these values as a matter of fact.
> 
> Pavel.
> 
>> we've solved a large structure (~20,000 residues/asymm unit), with 4-fold ncs and diffraction data to 3.3A.
>>
>> The Rfree/Rfac is ~28%-24% with OK geometry with no major outliers in the Ramachandran plot.
>> I would think I'm done (.. after 6 years!). 
>> However, my refined model b-factor (~130A2) is >> Wilson b-factor (~80A2). Obviously I'm not too happy with it.
>>
>> Here is what I tried to resolve this discrepancy: --> play with wxu_scale --> play with B-factor weight in ncs restraint (4-fold ncs)
>> --> play with number of macrocycles --> Redefine tls groups
>>
>> So far nothing really works, except switching from
>> individual_adp to group_adp. However, this increases my Rfree by almost 3%.
>>
>> Any ideas?
>>  
> _______________________________________________
> phenixbb mailing list
> phenixbb at phenix-online.org
> http://phenix-online.org/mailman/listinfo/phenixbb
> 
> 
> _______________________________________________
> phenixbb mailing list
> phenixbb at phenix-online.org
> http://phenix-online.org/mailman/listinfo/phenixbb