At the risk of derailing the discussion, can it be that the blob is actually an 
accumulation of many Fourier ripples? (on top of bulk solvent, I guess). The 
“chloride” seems to be about 3.5Å away from a lot of atoms, with nothing closer.
This is mostly based on intuition and the fact that in my experience any almost 
spherical cavity or any almost cylindrical crevice has a blob of difference 
density inside, which often proves to be very difficult to model. I have no 
hard data to back this up.

Cheers,

Jose.

================================
Jose Antonio Cuesta-Seijo, PhD
Carlsberg Laboratory
Gamle Carlsberg Vej 10
DK-1799 Copenhagen V
Denmark

Tlf +45 3327 5332
Email 
[email protected]<mailto:[email protected]>
================================

From: CCP4 bulletin board [mailto:[email protected]] On Behalf Of Bernhard 
Rupp (Hofkristallrat a.D.)
Sent: Thursday, January 22, 2015 10:26 AM
To: [email protected]
Subject: Re: [ccp4bb] chloride or water

After reading this exchange, I think at the core of the dispute is the question 
what a structure
model really is supposed to represent (a), and how to annotate/describe it (b).

ad (a) In general, and forgive me for not disclosing all caveats and fine  tune 
(I leave this to GB), we are
interested in the posterior likelihood (model likelihood). The two terms to 
consider (yes, I know, I am omitting
any normalization necessary for hypothesis testing etc) this model likelihood 
would be proportional to
the product of an evidence term (data likelihood) and an independent prior 
knowledge term.

Imho the expressed opinions diverge primarily in the relative significance of 
the terms or normalization of the probabilities.
The evidence purists (and it seems that computationalists often mistake this 
for arrogance of the crystallographers)
argue that if I can’t see/recognize it in ED or support it otherwise by direct 
experimental evidence, leave it out
of the model (after all, X-ray structure models are supposed to be based on 
experimental evidence).

On the other hand, from prior knowledge (admittedly extracted from polluted 
data bases like the PDB
and that is not meant as an insult but a statement of fact) we do know 
something about what reasonably could be
expected and could use it to the full extent of its statistical support.

Both extremes are of course justifiable, but in practice not separable. E.g. we 
use riding hydrogens without giving it a
second thought that we do not see them in (macro X-ray) ED, and they do improve 
models. On the other hand,
we still put side chain atoms we do not ‘see’ in specific positions and hope 
that the B-factors increase to a point where
the absence of any meaningful scattering contributions does not ruin our Holy 
R. That specific position is perhaps
closer to ‘wild speculation’ than the probability that a chloride atom exists 
in that specific case.

(I do argue that in the above case a set of conformations with occupancies of 
rotamers corresponding to their
population in the torsion angle landscape (or in the polluted databases) – the 
prior – under consideration where they
cannot be – the rest of the model as obtained from evidence – would be a 
possible description).

The final weighting one could apply might be a less tangible factor – how badly 
does it matter? If a ligand in a specific
pose is modelled and intended for the use of drug discovery, I’d say the claim 
is extraordinarily strong, and
the model likelihood (both terms) better be convincing. In the less earth 
shaking blob case, considering priors and the mentioned
restrictions of low resolution etc, I can accept a low but not unreasonable 
probability (<- such apparent evasiveness being a dead
giveaway of a mental Bayes factor calculation instead of adherence to an 
artificial significance level; frequentists please feel free
to flame me) for Cl as the most probable in the Cl/water/empty model 
competition (not that any of the models are
overly convincing, however, compatible with the low drama factor of that 
decision).

ad (b) having said this, how to express such probabilistic considerations in 
the current atomic PDB model format, is an unresolved
issue. I think the whole idea of the single static atomic model sooner or later 
will fall. It is already a mess because
much information about the model is hidden for example in remarks like TLS 
groups (btw, one of the most abused and
ad-hoc applied means in the hope of reducing Holy R instead of reflecting what 
these groups actually mean). But this is besides the
original point and becoming free floating…

I am not calling for making peace here, rather argue that the seemingly 
insignificant issue of a single Cl ion in one of
100k structure models can lead to productive reflection about meaning and 
improvement of model description.
Sorry for offending those in need for cozy comfort closing quotes.

The answer is, as always, 42.

HTC, BR

(Happy To Confuse)



From: CCP4 bulletin board [mailto:[email protected]] On Behalf Of Keller, 
Jacob
Sent: Mittwoch, 21. Januar 2015 19:18
To: [email protected]<mailto:[email protected]>
Subject: Re: [ccp4bb] chloride or water

I reiterate that assigning a chloride is not “wild speculation” or “just making 
something up” in light of what we know about the situation.

I see your point about not knowing that it’s a chloride, but I think you would 
agree that it is certainly more likely a chloride than map-noise, and perhaps 
more likely than water as well. Would you agree that chloride is the best 
guess, at least? What are the options for that blob, and what is the 
probability of each? I think you want to make sure people don’t get misled by 
it, which is a good point and a noble aspiration. I would argue that “not 
choosing” is here, as everywhere else, indeed choosing. And if you choose 
nothing here, you are almost certainly wrong, given the data. Some might be 
surprised or misled that a cavity like that would be totally empty, and the map 
density is unequivocal evidence that something is indeed there. So what now? 
Maybe a solution would be dummy atoms, maybe call them agnostons (agn) or 
something? Perhaps this is a basic disagreement about what a protein structure 
model represents, with one opinion being “that which we can rely upon 
confidently” and the other being “that which is most likely considering the 
data.” Each has advantages depending on one’s goals. Since the PDB is certainly 
tainted by structures modeled in accordance with the “most likely” outlook, one 
now has to be cautious about all structures.

I prefer the latter since I would try to be responsible/skeptical enough to go 
back to the original crystal data before making important conclusions based on 
it. Maybe there should be a disclaimer printed in each PDB file…

JPK


From: CCP4 bulletin board [mailto:[email protected]] On Behalf Of Nat Echols
Sent: Wednesday, January 21, 2015 12:33 PM
To: [email protected]<mailto:[email protected]>
Subject: Re: [ccp4bb] chloride or water

On Wed, Jan 21, 2015 at 9:05 AM, Keller, Jacob 
<[email protected]<mailto:[email protected]>> wrote:
Not sure why there is this level of suspicion about the poor halide when waters 
generally get assigned so haphazardly. I would say that there are probably more 
“wrong” waters in the PDB than wrong chlorides, but there’s not much fuss about 
that.

Great, so leave it empty instead of just making something up.  Perhaps future 
generations will figure out a more rigorous and quantitative method for 
handling such features than guessing based on screenshots posted to a mailing 
list.  At this resolution water placement is difficult to justify anyway - and 
since neither the scattering properties nor the coordination distances are 
especially accurate, trying to assign chemical identity in the absence of any 
supporting information (for example anomalous data) is especially futile.
(Although at least in this case the resolution is an obvious red flag - to a 
crystallographer, anyway - indicating that any lighter ions shouldn't be taken 
very seriously.  Other biologists, of course, may be more trusting.)
-Nat

Reply via email to