[phenixbb] individual anisotropic adps

Pavel Afonine PAfonine at lbl.gov
Thu Dec 11 11:57:08 PST 2008

Hi Frank,

- I tend to call it "isotropic atomic displacement parameters" and 
"anisotropic atomic displacement parameters", correspondingly "isotropic 
ADP" and "anisotropic ADP", which is also consistent with:
J. Appl. Cryst. (2002). 35, 477-480   
On the handling of atomic anisotropic displacement parameters
R. W. Grosse-Kunstleve and P. D. Adams

- What you wrote below can be boiled down to the question: "Are very 
tight restraints equivalent to constraints?". A few years ago I tried to 
see if refinement of individual coordinates with very tight restraints 
can approach rigid body refinement. The answer I got was "no". Of course 
one needs to keep in mind that it might all be dependent on particular 
implementation, minimizers, and other technicalities. I don't know what 
mathematics says here.

- I would just take whole PDB and re-refine it alternative options at 
appropriate resolutions (for example, use iso- or anisotropic ADP for 
model at resolution between 1.2 and 2.0A) and see how the best choice 
correlates with model/data characteristics. I'm sure there will be 
dependencies allowing you to get some empirical rules. However, I'm sure 
there will be a number of exceptions, where only trying appropriate 
options will bring you the definitive answer.

- Using TLS is another branch of this story. The total atomic B-factors 
can be presented at least as a sum of three contributions:

Utotal = Ulocal + Utls + Ucryst

    - Utotal is the total ADP,
    - Ulocal reflects the local atomic vibration (also named as residual 
B) and should obey Hirshfeld's rigid bond criteria (F.L. Hirshfeld. Acta 
Cryst. (1976). A32, 239-244. "Can X-ray data distinguish bonding effects 
from vibrational smearing?");
    - Ucryst reflects global lattice vibrations (Ucryst is determined 
and refined at anisotropic scaling),
    - Utls reflects global motion (which can be further subdivided into 
the motion of the molecule as a whole and into the motion of its domains).

Obviously, the presence of global motion is not the function of data 
resolution (in sense that a molecule does not "know" about your 
crystallographic experiment and its resolution), so ideally you need 
always model it (= you need always use TLS). However this faces a number 
of technical problems currently making it impossible:
- robust and reliable choice of TLS groups;
- robust and reliable separation of  Utls and *anisotropic* Ulocal. A 
first attempt is described here: P. Afonine & A. Urzhumtsev. (2007). 
CCP4 Newsletter on Protein Crystallography. 45. Contribution 6. "On 
determination of T matrix in TLS modeling";
- not to mention a strong correlations of parameters which potentially 
may lead to troubles in optimization flow;
- ... and many other...

Phenix.refine has a pretty sophisticated algorithm of combined TLS 
refinement which is outlined at slides #28-32:


On 12/11/2008 9:08 AM, Frank von Delft wrote:
> Question at large:  do any formalisms exist, or are any envisioned, that 
> deal with this question more rigorously?  E.g. in principle everything 
> is always ADP, but as resolution decreases, the restraints are tightened 
> accordingly, so that there is an appropriately refined continuum from 
> fully ADP @ 1A to isotropic @ 2A to TLS only @ 3A?
> I know for a fact I've heard about this somewhere, but can't remember where.
> phx.
> that in principle always refine ADP for everything, but tighten the 
> restraints automatically,
> Pavel Afonine wrote:
>> Hi Gerwald,
>> from what I recall is if you refine anisotropic ADP at "too low" 
>> resolution in other refinement programs, they were often stopping 
>> complaining about non-positive definite U or unstable refinement. This 
>> seemed to me the main technical limit.
>> In phenix.refine this does not happen but it doesn't mean that refining 
>> anisotropic ADP at "too low" resolution will yield meaningful results.
>> I'm not sure that there is an exact resolution borderline for doing (or 
>> not doing) individual anisotropic ADP refinement. From my experience, 
>> I've never seen a case at ~1.5A and higher where switching from iso- to 
>> anisotropic model was not good. Then this question transforms into "at 
>> which resolution you start refining waters as anisotropic?"... A very 
>> rough number I would say is 1.2-1.1A and higher.
>> In-between 1.5 and 1.7 (1.8A) is a "gray zone", where it is very likely 
>> that anisotropic ADP refinement will result in over-fitting, however 
>> there is a chance that it might still be valid.
>> Summarizing, if I have a data at resolution between between 1.5 and 1.8A 
>> resolution, I just try two refinements: in one refining isotropic and in 
>> another refining anisotropic ADPs. Rfree is your best friend here (and 
>> Rfree-Rwork). In my opinion, this is the most robust and cheep way to 
>> get the answer.
>> Cheers,
>> Pavel.
>> On 12/11/2008 7:48 AM, gerwald jogl wrote:
>>> Hi All,
>>> I was wondering if there is information/recommendations out there for 
>>> the resolution required to refine individual anisotropic adps. I think 
>>> the recommendation for shelx is at least 1.5 A resolution.
>>> In addition, is phenix set up and would it make sense to combine tls 
>>> with individiual anisotropic adps?
>>> All comments welcome,
>>> Gerwald
>>> _______________________________________________
>>> phenixbb mailing list
>>> phenixbb at phenix-online.org
>>> http://www.phenix-online.org/mailman/listinfo/phenixbb
>> _______________________________________________
>> phenixbb mailing list
>> phenixbb at phenix-online.org
>> http://www.phenix-online.org/mailman/listinfo/phenixbb
> _______________________________________________
> phenixbb mailing list
> phenixbb at phenix-online.org
> http://www.phenix-online.org/mailman/listinfo/phenixbb
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://phenix-online.org/pipermail/phenixbb/attachments/20081211/cda02760/attachment-0003.htm>

More information about the phenixbb mailing list