Dear Ethan,
To move away from an abstract discussion of hydrogen atoms I'd like
to describe a concrete example. In 2008 I deposited a model of the FMO
(Bacteriochlorophyll containing) protein. The ID code is 3EOJ. The
model was refined to a data set cut off at 1.3 A resolution using the
criteria of the day. I used shelxl for the final stage of refinement
and added riding hydrogen atoms to the mix. When I deposited the model
I succumb to peer pressure and removed the hydrogen atoms.
If you look at the map calculate by the Electron Density Server you
will see many peaks in the Fo-Fc map indicating the missing hydrogen
atoms. (I have attached a screen-shot from Coot but I recommend that
you fire up Coot and explore the map yourself.) In my picture you can
see the three peaks around a methyl group. Above and to the left is the
peak for the hydrogen of a CH bridging atom in the Bacteriochlorophyll-a
ring. To the right and in the distance is a peak for the hydrogen of a
CH2 group. Not every hydrogen is represented by a positive peak, but
they exist throughout the map. This Fo-Fc map is useless for the
purpose of assessing the quality of my model, since the true residuals
are hidden among all these hydrogen peaks.
A critic might say that these peaks are simply the result of the
model being biased toward the presence of hydrogen atoms and therefore
an artifact. A model refined to this data set w/o hydrogen atoms would
not likely show peaks indicating that hydrogen atoms need to be built.
I would say that the map calculated from a Hydrogen-free model is the
biased one. I am 99% confident in the location of most of the riding
hydrogen atoms and leaving them out results in a model that is
fantastically unlikely. The absence of peaks in an apo map is a flaw in
that map. I would describe it as "vacuum bias". "Biasing" a model
toward reality is not a problem.
This example shows that the current PDB is incompatible with models
whose Hydrogen atoms have been stripped. To get proper maps and
validation reports one has to either preserve the Hydrogen atoms in the
model, or modify all the software that uses coordinate files to add the
hydrogen atoms back in. That is a major programming task, which the
authors have, apparently, been unwilling to do.
I will go further and disagree with you that even this is a solution.
It is very difficult to add the Hydrogen atoms back into 3EOJ, and I
expect this difficulty is the reason software has not been written that
successfully does it.
There are two major problems to be overcome in 3EOJ. shelxl has an
option to twirl the methyl groups and select the torsion angle with the
best fit to the map. The hydrogen atoms in the pictured methyl group
weren't built as staggered -- All values for the torsion angle were
tested and it happens that the best fit placed them in a staggered
conformation. That is a much more interesting result. There are other
methyl groups around the edges of the Bchl-a molecules that are crowded
and the methyl groups are observed to have torsion angles that are not
standard for riding Hydrogen atoms. The neighboring methyl groups avoid
H-H bumps by twisting and that twist can be detected by shelxl in the
1.3 A data.
The second problem is the matter of Histidine residues. There are
two Nitrogen atoms in the side chain. A hydrogen atom could be on
either one, and sometimes both have hydrogens. A very clever program
could work out from the hydrogen bonding pattern the most likely
placement, but I've not seen any program that is very good with hydrogen
bonding networks. Worst still, I've often seen programs place the
hydrogen atom *between* the Nitrogen and Magnesium atoms of a Histidine
ligand to a Bacteriochlorophyll a. This mistake will certainly lead to
very bad geometry!
Until an hydrogenation program is written that can handle all
ligands, all hydrogen bonding networks (even overlapping partially
occupied ones), and use the imaging data to place the Hydrogen atoms
that have one or two dimensional mobility, I don't see an alternative to
leaving the (hopefully) manually inspected and curated Hydrogen atoms in
the deposited PDB.
Dale Tronrud
P.S.
I'm hoping to find the time to use the new versioning capability of
the wwPDB to put my Hydrogen atoms back in 3EOJ.
On 2/28/2020 10:34 PM, Ethan A Merritt wrote:
> Matthew:
>
> I think your nice summary leaves out an important point that has not been
> explicitly mentioned. That is the question of whether depositing hydrogens
> actually adds information to the model. I submit that for a typical protein
> refinement it does not. The model is adequately described by saying
> "hydrogens were added in their riding positions". This, together with
> knowledge of the refinement program used, is sufficient to reconstruct
> the full model.
>
> This is an example of a recurring concern of mine that model validation
> should include consideration of whether the model is overly complex.
> Unless you have an abundance of data (which admittedly your 1.0Å case
> might) there are insufficient observations to refine 3 positional parameters
> for each hydrogen as if they were free variables. We typically bypass this
> by instead using the riding hydrogen model, which adds effectively a single
> on/off parameter for the entire mode (plus a small number of implicit
> parameters that describe the ideal riding geometry, but those are
> normally taken as a priori knowledge rather than free variables).
>
> So I find deposition of hydrogens for a typical resolution structure to
> be more misleading than useful. The correct, parsimonious, description is the
> one-line statement that a riding hydrogen model was used.
>
> It is tangential to your question, but I hold the same view about depositing
> ANISOU records for a structure when the source of the anisotropy is solely
> a TLS model, either with or without individual Biso contributions.
> The parsimonious description is to give the TLS parameters and the Biso
> component, if any. These can be expanded to regenerate per-atom
> ANISOU parameters if desired by a downstream program.
> If you deposit ANISOU records it implies that the Uij terms they describe
> are free variables, but they are not. (or anyway IMHO they should not be,
> although PHENIX can violate this stricture).
>
> My view is that for a typical structure (i.e. worse than say 1Å resolution data)
> depositing hydrogen positions and ANISOU records at best does no harm.
> Unfortunately it implies a statistically unjustifiable model treatment.
> The justifiable model is adequately described by the small number of
> parameters in the header records; the hydrogen coordinates and ANISOU
> parameters are redundant dross.
>
> I fully understand that your original question was driven by cases where
> you do have very high resolution data and so the statistical justification of
> refining individual hydrogens or anisotropic ADPs enters a different realm.
>
> Ethan
>
>
> On Friday, 28 February 2020 20:22:17 PST Whitley, Matthew J wrote:
>> Dear all,
>>
>> I want to thank everyone who responded to my query about whether or not to
>> include hydrogens in PDB depositions when they were explicitly included in
>> the model during refinement. In addition to the replies posted to this
>> bulletin board, I received numerous replies sent directly to my email
>> address.
>>
>> To clarify one more time for casual readers so that we are all on the same
>> page: because these two structures happen to be at high resolution (1.0 and
>> 1.2 Å, respectively), I decided to include explicit hydrogens in the model
>> for refinement, as recommended by the documentation for both Phenix and
>> Buster, which I used for these refinements. For the Phenix refinements,
>> hydrogens were added by phenix.ready_set, whereas for the Buster
>> refinements the hydrogenate tool was used. My understanding is that both
>> of these eventually call the reduce tool from MolProbity. Unsurprisingly,
>> the presence of hydrogens on the model led to both better model geometry
>> and lower R-factors, although at these resolutions there is no observable
>> density for the vast majority of the H-atoms in any of the refined maps.
>>
>> Because the presence of the hydrogens improved the model, I have decided to
>> leave the hydrogens at their refined positions for deposition.
>>
>> I do want to point out one thing for readers interested in this topic: based
>> on all the replies I received, there are a number of differing opinions
>> (and therefore different practices) as to whether hydrogens should be
>> included in deposited structures. The expressed opinions ranged from the
>> ethical (if the hydrogens were there for refinement, then it’s only fair
>> that they be present in the deposited structure so that downstream users
>> know what went into generating the reported statistics) to the practical
>> (if the paper’s conclusions don’t rely on any arguments based on hydrogen
>> atom positions, then there’s no compelling reason for them to be there;
>> include them or don’t, it doesn’t matter.) Because opinions seem to vary,
>> perhaps it would be worthwhile for the PDB to issue some guidance on the
>> matter for the future.
>>
>> Have a nice weekend, everyone.
>>
>> Matthew
>>
>> ---
>> Matthew J. Whitley, Ph.D.
>> Research Instructor
>> Department of Pharmacology & Chemical Biology
>> University of Pittsburgh School of Medicine
>>
>
> ########################################################################
>
> To unsubscribe from the CCP4BB list, click the following link:
> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCP4BB&A=1
>
########################################################################
To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCP4BB&A=1
|
