How about this one:
Can't
Refine
Your
Structure
To
Acceptable
Likelihood
Typical error bars on crystallographic data are ~5% (R_sym), but with
very few exceptions the models in the PDB do not fit their corresponding
observations to better than ~40% error (R_cryst for intensities).
Gerard? What is the "likelihood" that such a model is correct?
Jokes and jabs aside, how much can we trust conclusions based on a model
with such a large amount of systematic error? I have been looking for
the answer to this question for many years now. So far, no luck.
DePriso et al. (2004) Structure 12, 831-8 framed this problem much
better than I just did. In my experience, pretty much everyone has a
hypothesis of why crystallographic R factors are so high: multiple
conformers is a popular one, as is semi-ordered solvent, local minima in
refinement space, etc. but I have yet to find convincing experimental
evidence (in the form of a 5% R_cryst with observations/parameters > 1
as is generally required of small molecule structures), or even a
controlled experiment to verify or reject any of these hypotheses. For
example, Vitkup et al. showed that fitting a single model to
MD-simulation derived "data" gave ~20% R, which means multiple
conformers are sufficient to explain the "R-Factor Gap", but the
converse has never been shown. The best results I have seen modeling
multiple conformers (such as the seminal Burling et al. 1996) get a few
percent decrease in R, but nothing close to the 15% needed to close the
"R-Factor Gap". Anybody got an idea for a necessary AND sufficient test?
I know that the core reason why we believe that our models are correct
is because they visually agree with the "1-sigma" contour of
experimentally-phased electron density maps. But, when it comes to
comparing NMR and MX, I am reminded of a certain idiom ... involving
glass houses.
-James Holton
MAD Scientist (who has also done a little NMR).
Gerard Bricogne wrote:
> Dear Tassos, Bernhard and David,
>
> If I may push this humourous response (obviously tainted with
> crystallographic bias) a little further, I would say that my favourite
> mnemonic for the acronym "NMR" is
>
> N eeds
>
> M ore
>
> R esolution
>
> Joking apart, of course, it is a devilishly clever method.
>
>
> With best wishes,
>
> Gerard.
>
>
> --
> On Fri, Nov 14, 2008 at 11:28:25AM +0100, Anastassis Perrakis wrote:
>
>> Since I don't like attachments, I will first iterate the title of the
>> attached publication:
>>
>> "Traditional Biomolecular Structure Determination by NMR Spectroscopy
>> Allows for Major Errors "
>>
>> It immediately reminded me of an older one (ehm .. one author in common!),
>> addressed at that time mostly to crystallographers:
>>
>> "Errors in protein structures." (Nature 1996)
>>
>> ... and I am afraid that the authors were right in both cases (they did not
>> make many friends publishing these though)
>>
>> Crystallographers learned from that paper back then. And the participation
>> of NMR spectroscopists on the 2006 paper
>> implies they are also learning ;-)
>>
>> A.
>>
>> On Nov 14, 2008, at 6:34, Bernhard Rupp wrote:
>>
>>
>>>> wondering what people think of this.
>>>>
>>> Very funny.
>>>
>>> But no kidding, Richard Dickerson, the pioneer of DNA crystallography,
>>> comes from your institution. For DNA, NMR has the benefit of readily
>>> identifying intercalations in short oligomers etc w/o agony of
>>> crystallizing.
>>>
>>> For others, pls see attached. As a physical principle, spectroscopic
>>> methods do not deliver atomic resolution structures, but a set of
>>> inferences that may or may no be compatible with a molecular model.
>>>
>>> BR
>>> <Nabuurs_2006_PLOS_biomolecular_structure_NMR_errors.pdf>
>>>
>
>
|