[phenixbb] Refinement of B-factors at low resolution

Pavel Afonine pafonine at lbl.gov
Sat Mar 23 09:10:47 PDT 2013

Hi Florian,

Wilson B is just an estimate, and your refined B reflect your data.. 
Assumptions made to derive Wilson B include uniformly distributed atoms 
in the unit cell, same atom type and B-factor (what else?). This is not 
quite true for protein crystals, which typically consist of atoms of 
different kind, have different degree of disorder (B-factors), and they 
kind of form some structure (definitely not randomly distributed in the 
unit cell). So to me it's a miracle that oftentimes Wilson B closely 
matches refined average B. We should acknowledge that when it doesn't 
that's a valid possibility.

Regarding papers.. Exactly that plot I sent you before was not published 
(I guess it did not go to the phenixbb for the reason Nat explained 
yesterday), but the tools and methods are described here:

Afonine, P.V., Grosse-Kunstleve, R.W., Chen, V.B.,Headd, J.J., Moriarty, 
N.W., Richardson, J.S., Richardson, D.C.,Urzhumtsev, A.,Zwart, P.H. & 
Adams, P.D. (2010).*phenix.model**_vs_data**: a high-level tool for the 
calculation of crystallographic model and data statistics*. J. 
Appl.Crystallogr. 43, 669-676.

Urzhumtsev, A., Afonine, P.V. & Adams P.D. (2009).*On the use of 
logarithmic scales for analysis of diffraction data*.ActaCryst. D65, 

Urzhumtseva, L., Afonine, P.V., Adams, P.D. &Urzhumtsev, A. 
(2009).*Crystallographic model quality at a glance*.ActaCryst. D65, 297-300.

and the source of information used for all the analysis described in 
above papers is PDB. Also, I guess, you can obtain that plot from PHENIX 


On 3/23/13 6:53 AM, Florian Brückner wrote:
> Thanks, Pavel.
> So you don't think that the difference between Wilson B (66.45 
> according to phenix) and average B (52 after refinement) is still too 
> large? Where did you take the statistics from? Is that published anywhere?
> Best regards
> Florian
> Am 22.03.2013 um 07:09 schrieb Pavel Afonine <pafonine at lbl.gov 
> <mailto:pafonine at lbl.gov>>:
>> Hi Florian,
>> I think it is pretty convincing that refinement went all right and 
>> the mean B you get is what it actually is. Also I just looked at the 
>> files that you sent me the other day and I did not notice anything 
>> suspicious.
>> Pavel
>> P.S.: Sorry if you received multiple copies of this email: I was 
>> trying to attach a tiny image with some PDB statistics but our server 
>> did not let me do it.
>> On 3/21/13 1:00 AM, Florian Brückner wrote:
>>> Dear Pavel
>>> yes, the refinement did converge. I did 6 macrocycles and in 
>>> separate runs started from different average atomic B-factors of 32, 
>>> 66 and 97 keeping the distribution as in the original high 
>>> resolution molecular replacement model from the same crystal form 
>>> (i.e. adding 0, 34 and 65 to all atomic B-factors in the starting 
>>> pdb). In all cases after 6 macrocycles I get an almost identical 
>>> distribution of B-factors and also average B-factor of about 52. 
>>> Also setting all B-factors to the same value of 66 gave a similar 
>>> distribution after refinement with an average of 52. However, in the 
>>> latter case the R-free was about 2% higher compared to starting from 
>>> the original distribution, but it dropped by about 1% during refinement.
>>> Best regards
>>> Florian
>>> Am 21.03.2013 um 06:40 schrieb Pavel Afonine <pafonine at lbl.gov 
>>> <mailto:pafonine at lbl.gov>>:
>>>> Hi Florian,
>>>> I'm glad it works now. Regarding difference between Wilson B and 
>>>> mean refined B: well, it's hard to quantify what's "too low" or 
>>>> "too different". Average refined B may deviate from Wilson B and 
>>>> the difference in your case does not seem to be terribly large.
>>>> Did the refinement converge? What if you do say 10 or 20 
>>>> macro-cycles instead of default 3? Will B-factors further change 
>>>> between macro-cycles?
>>>> Pavel
>>>> On 3/19/13 7:37 AM, Florian Brückner wrote:
>>>>> Dear Keitaro and Pavel,
>>>>> thank you for your suggestions. I have now used the latest nightly 
>>>>> build (1326). With this version, the average B-factor converges at 
>>>>> about 52 A**2 and also the distribution is very similar, 
>>>>> regardless with what B-factors I start the refinement (I tried 
>>>>> average B-factors between 32 and 97). So clearly the behavior of 
>>>>> this newest version is different from the version I have been 
>>>>> using before (1.8.1 stable release 1168) and also from the latest 
>>>>> stable release (1.8.2 stable release 1309), which behaves the same 
>>>>> as 1168. To me the B-factors still look a bit too low, since the 
>>>>> Wilson B-factor calculated by phenix is 66.45, the Wilson B-factor 
>>>>> I determined manually from the linear part of the Wilson plot is 97.
>>>>> Best regards
>>>>> Florian
>>>>> Am 19.03.2013 um 00:06 schrieb Keitaro Yamashita 
>>>>> <yamashita at castor.sci.hokudai.ac.jp 
>>>>> <mailto:yamashita at castor.sci.hokudai.ac.jp>>:
>>>>>> Dear Florian,
>>>>>> Maybe you are using phenix-dev-1311 or earlier?
>>>>>> I am afraid that you have experienced the bug that overall 
>>>>>> B-factor is
>>>>>> not applied to individual atomic B-factors in certain cases, which I
>>>>>> have also experienced recently. That bug was fixed in dev-1312 if I
>>>>>> recall correctly.
>>>>>> Hope it helps,
>>>>>> Keitaro
>>>>>> 2013/3/19 Florian Brückner <florian.brueckner at psi.ch 
>>>>>> <mailto:florian.brueckner at psi.ch>>:
>>>>>>> Dear all
>>>>>>> I am refining a structure at 3.2 A with phenix.refine using 
>>>>>>> automatic
>>>>>>> optimization of target weights, solved by molecular replacement 
>>>>>>> with a 2.1A
>>>>>>> structure of the same crystal form. Rwork=26%, Rfree=28%. I use 
>>>>>>> individual
>>>>>>> atomic B-factor refinement. The obtained distribution of 
>>>>>>> B-factors after
>>>>>>> refinement looks reasonable and quite similar to the 
>>>>>>> distribution in the
>>>>>>> higher resolution model, which I suppose would be expected, 
>>>>>>> since it is the
>>>>>>> same crystal form. However, the average B-factor does not change
>>>>>>> substantially, i.e. if I start with the B-factors from the high 
>>>>>>> resolution
>>>>>>> model (average B-factor around 32 A**2) after 6 macrocycles the 
>>>>>>> average
>>>>>>> B-factor is still around 32 A**2. If I set all B-factors to the 
>>>>>>> Wilson
>>>>>>> B-factor determined by phenix.refine (66 A**2), after 6 
>>>>>>> macrocycles the
>>>>>>> distribution of B-factors looks again very similar to the high 
>>>>>>> resolution MR
>>>>>>> model, but the average B-factor is still around 65 A**2. What 
>>>>>>> should I make
>>>>>>> of this? Is this an expected behavior at this resolution, i.e. 
>>>>>>> the fact that
>>>>>>> the absolute values of B-factors cannot be determined during 
>>>>>>> refinement,
>>>>>>> although a reasonable distribution can be obtained as judged by 
>>>>>>> plotting
>>>>>>> average B-factors vs. residue#? Since the average B-factor does 
>>>>>>> not change
>>>>>>> much during refinement, which average B-factor should I start 
>>>>>>> refinement
>>>>>>> with to end up with reasonable absolute values? I have manually 
>>>>>>> determined
>>>>>>> the Wilson B-factor from the linear part of the Wilson plot 
>>>>>>> (4.5-3.2 A) to
>>>>>>> be 97 A**2. This seems high to me. Would that be expected for 
>>>>>>> the given
>>>>>>> resolution, given that the 2.1A structure of the same crystal 
>>>>>>> form has 32
>>>>>>> A**2? I should add, that there is not a substantial effect of 
>>>>>>> the average
>>>>>>> refined B-factor on Rfree.
>>>>>>> Thanks for any comments or suggestions.
>>>>>>> Best regards
>>>>>>> Florian

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://phenix-online.org/pipermail/phenixbb/attachments/20130323/17e46b7e/attachment.htm>

More information about the phenixbb mailing list