[phenixbb] Persistant high average B-factor.

[Pavel Afonine]

Hi Everyone,

just this morning I replied someone else offlist to a similar question. 
I replied to Joe offlist too. I guess I had to actually reply to 
phenixbb, so I wouldn't have to duplicate emails multiple time...

This is a very typical question and honestly, and sometime I feel like 
people get too sensitive to "high" B-factors...

The atomic displacement parameter tends to measure several physical 
phenomena, including: 1) dynamic disorder, that is how much an atom 
moves around its position defined by the atomic model, and 2) static 
disorder arising from (typically small) differences in equivalent atomic 
positions throughout the whole crystal. In addition, it measures the 
correctness of atomic model – a misplaced atom would have outstandingly 
low or high value of refined ADP compared to surrounding correctly 
placed atoms. Clearly this is way more than the mathematical model 
implemented for it in refinement is supposed to model.

Within structure refinement framework the B-factors are supposed to 
model only the first phenomenon (dynamic disorder), while the others 
effects get absorbed by irrelevant refinable parameters contributing to 
their uncertainties. Large-scale static disorder can be modeled, for 
example, by alternative conformations or by using anharmonic functions 
for ADP (Acta Cryst. (2009). D65, 284-293) or multiple models.

Having said that, the refined value of a B-factor is not necessarily 
directly related to mean amplitude of harmonic vibration of an atom. The 
matter of fact is that it has much more in it. So when you get a huge 
B-factor I wouldn't take it too literally and convert it into angstroms 
of mean amplitude of its harmonic vibration.

Sometime people believe that their refined B-factors should closely 
match the Wilson B, and they get very excited if it not the case and 
start talking about "a big difference between mean B and Wilson B". But, 
they probably forget one "tiny" detail, that:

- the Wilson B-factor is just an estimation based on assumption that the 
atoms are randomly distributed in the unit cell. Clearly the atoms in 
your model are not randomly filling the unit cell! Plus, the accuracy of 
this estimation drops with the resolution.

- refined B-factors is what reflects your data, so I would trust these 
values more rather than an estimate obtained under unrealistic assumptions.

Given this, can anyone seriously tell what is "big difference" (see 
recent Dale Tronrud's great reply on this matter)?

Now, if we look at thousands of structures in PDB to see how their mean 
B-factors are distributed as a function of resolution, we will get 
something like this:

Resolution, A <B>
1.25-1.5 18.96
1.5-1.75 21.99
1.75-2.0 26.51
2.0-2.25 31.75
2.25-2.5 38.53
2.5-2.75 42.49
2.75-3.0 48.01
3.0-3.25 55.8
3.25-3.5 65.25

Joe refines his structure at 1.7 and his mean B~40A**2, and the Wilson B 
is ~30. Yes, both refined and Wilson B are somewhat larger than observed 
in average over PDB structures. This fact would prompt me to 
triple-check if the correct refinement strategy is used, and if 
everything else looks ok, and if everything is good then I would not 
worry too much about the B-factors - there must be something in your 
crystal/data that they (B-factors) are trying to convey to you.

All the best!
Pavel.


On 6/10/10 1:02 PM, Nian Huang wrote:
> Hi Joseph,
> I had a similar situation. In the end, I ended up stopping using TLS
> for my structure. Do you have the same problem by just doing
> individual adp refinement?
>
> Nian
>
> On Mon, Jun 7, 2010 at 8:52 PM, Joseph Brock <joeylives2ride at hotmail.com> wrote:
>   
>> Dear Phenix users,
>>
>> I am currently using version 1.6.1-353 and I'm working on a 1.7A structure
>> that has refined quite well using
>> Individual_sites+individual_adp(isotropic)+occupanices+TLS: Final R-work =
>> 0.1776, R-free = 0.2262
>>
>> However, I noticed that the average B-factor reported by polygon with these
>> settings was quite high ( < 0.1 of other structures with a similar
>> resolution).
>>
>> Discovering I could not change the value of wxu_scale when using
>> individual_adp+TLS, I went about trying to reduce the average B using the
>> method described here:
>>
>> http://phenix-online.org/pipermail/phenixbb/2007-September/000511.html
>>
>> However,  I still can't seem to reduce the average B-factor. I've just done
>> a refinement with the following non-default (extreme) target weight
>> settings:
>>
>>   target_weights {
>>     wxc_scale = 1.5
>>     fix_wxc = 2
>>     fix_wxu = 0.1
>>    (wxu_scale = 1)
>>
>> But if anything, reducing wxu to such low values makes the avergae B-even
>> higher while concuretnly making the gap between R-values much worse:
>>
>> Final R-work = 0.1602, R-free = 0.2457
>>
>> This suggests to me that I'm obviously over fitting the data but I'm at a
>> loss on how to proceed. Should i just accept that the automatic scaling used
>> for Phenix TLS wxu is doing a good job and that my structure actually does
>> have an unusually high avergae B?
>>
>> I'm fitting two ligands into density, which is at several places ambiguous,
>> so I would like the most unbiased difference density possible.
>>
>> Many thanks in advance,
>>
>> Joe.
>>