On 1/13/2022 11:14 AM, Tristan Croll wrote:
(please don’t actually do this)
Too late! I've been doing that for years. What happens, of course, is
the "geometry" improves, but the R factors go through the roof. This I
expect comes as no surprise to anyone who has played with the "weight"
parameters in refinement, but maybe it should? What is it about our
knowledge of chemical bond lengths, angles, and radii that is
inconsistent with the electron density of macromolecules, but not small
molecules? Why do macro-models have a burning desire to leap away from
the configuration we know they adopt in reality? If you zoom in on
those "bad clashes" individually, they don't look like something that is
supposed to happen. There is a LOT of energy stored up in those little
springs. I have a hard time thinking that's for real. The molecule is
no doubt doing something else and we're just not capturing it properly.
There is information to be had here, a lot of information.
This is why I too am looking for an all-encompassing "geometry score".
Right now I'm multiplying other scores together:
score = (1+Clashscore)*sin(worst_omega)*1./(1+worst_rama)*1/(1+worst_rota)
*Cbetadev*worst_nonbond*worst_bond*worst_angle*worst_dihedral*worst_chir*worst_plane
where things like worst_rama is the "%score" given to the worst
Ramachandran angle by phenix.ramalyze, and worst_bond is the largest
"residual" reported among all the bonds in the structure by molprobity
or phenix.geometry_minimization. For "worst_nonbond" I'm plugging the
observed and ideal distances into a Leonard-Jones6-12 potential to
convert it into an "energy" that is always positive.
With x-ray data in hand, I've been multiplying this whole thing by Rwork
and trying to find clever ways to minimize the product. Rfree is then,
as always, the cross-check.
Or does someone have a better idea?
-James Holton
MAD Scientist
On 1/13/2022 11:14 AM, Tristan Croll wrote:
Hard but not impossible - even when you *are* fitting to low-res
density. See
https://twitter.com/crolltristan/status/1381258326223290373?s=21 for
example - no Ramachandran outliers, 1.3% sidechain outliers,
clashscore of 2... yet multiple regions out of register by anywhere up
to 15 residues! I never publicly named the structure (although I did
share my rebuilt model with the authors), but the videos and images in
that thread should be enough to illustrate the scale of the problem.
And that was *with* a map to fit! Take away the map, and run some MD
energy minimisation (perhaps with added Ramachandran and rotamer
restraints), and I think it would be easy to get your model to fool
most “simple” validation metrics (please don’t actually do this). The
upshot is that I still think validation of predicted models in the
absence of at least moderate-resolution experimental data is still a
major challenge requiring very careful thought.
— Tristan
On 13 Jan 2022, at 18:41, James Holton <jmhol...@lbl.gov> wrote:
Agree with Pavel.
Something I think worth adding is a reminder that the MolProbity
score only looks at bad clashes, ramachandran and rotamer outliers.
MPscore=0.426∗ln(1+clashscore)+0.33∗ln(1+max(0,rota_out−1))+0.25∗ln(1+max(0,rama_iffy−2))+0.5
It pays no attention whatsoever to twisted peptide bonds, C-beta
deviations, and, for that matter, bond lengths and bond angles. If
you tweak your weights right you can get excellent MP scores, but
horrible "geometry" in the traditional bonds-and-angles sense. The
logic behind this kind of validation is that normally nonbonds and
torsions are much softer than bond and angle restraints and therefore
fertile ground for detecting problems. Thus far, I am not aware of
any "Grand Unified Score" that combines all geometric considerations,
but perhaps it is time for one?
Tristan's trivial solution aside, it is actually very hard to make
all the "geometry" ideal for a real-world fold, and especially
difficult to do without also screwing up the agreement with density
(R factor). I would argue that if you don't have an R factor then
you should get one, but I am interested in opinions about alternatives.
I.E. What if we could train an AI to predict Rfree by looking at the
coordinates?
-James Holton
MAD Scientist
On 12/21/2021 9:25 AM, Pavel Afonine wrote:
Hi Reza,
If you think about it this way... Validation is making sure that the
model makes sense, data make sense and model-to-data fit make sense,
then the answer to your question is obvious: in your case you do not
have experimental data (at least in a way we used to think of it)
and so then of these three validation items you only have one,
which, for example, means you don’t have to report things like
R-factors or completeness in high-resolution shell.
Really, the geometry of an alpha helix does not depend on how you
determined it: using X-rays or cryo-EM or something else! So, most
(if not all) model validation tools still apply.
Pavel
On Mon, Dec 20, 2021 at 8:10 AM Reza Khayat <rkha...@ccny.cuny.edu>
wrote:
Hi,
Can anyone suggest how to validate a predicted structure?
Something similar to wwPDB validation without the need for
refinement statistics. I realize this is a strange question
given that the geometry of the model is anticipated to be fine
if the structure was predicted by a server that minimizes the
geometry to improve its statistics. Nonetheless, the journal has
asked me for such a report. Thanks.
Best wishes,
Reza
Reza Khayat, PhD
Associate Professor
City College of New York
Department of Chemistry and Biochemistry
New York, NY 10031
------------------------------------------------------------------------
To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1
<https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1>
------------------------------------------------------------------------
To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1
<https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1>
------------------------------------------------------------------------
To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1
<https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1>
########################################################################
To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1
This message was issued to members of www.jiscmail.ac.uk/CCP4BB, a mailing list
hosted by www.jiscmail.ac.uk, terms & conditions are available at
https://www.jiscmail.ac.uk/policyandsecurity/
