Hi L?on,
On Mar 23, 2007, at 12:32 PM, L?on Salgado wrote:
>
> I did tried to run an anneal.py with and without RAMA. And the results
> clearly show that the X best structures from the annealing with
> RAMA do
> show a better agreement (rmsd from superposition) between the
> mainchain
> as well as the sidechains on the ensemble. Great!
You mean you've increased your protein structure's precision? That's
unusual. The
DELPHIC torsion angle potential (what I, too late, decided to call
the RAMA potential)
has minima corresponding to every observed combination of torsion
angles in the
database of xray structures it was built from. Thus, an
unconstrained sidechain won't
drop into a single conformation with refinement against the DELPHIC
torsions--it'll drop
into several conformations. The ensemble will be less of a
continuous smear of structures
at small scales (say, a single torsion angle), but it generally won't
look terribly different when
considering a protein structure as a whole.
That said, the DELPHIC torsion potential clearly results in more
physically-reasonable
structures, and Ad Bax showed several years ago that it results in
better agreement with
observed RDCs in the absence of direct refinement against those
RDCs. So there's no
question that DELPHIC torsions improve the accuracy of protein
structures. But it's still
rare to see any significant change in their precision.
>
> But I still have one question, about the rama stuff. When I was
> reading
> and trying to understand some scripts using rama
> (eg.eginputs/sry/sry_finall.inp), I saw that the torsional database
> has
> 3 sets of potentials: the raw (called with xrama), gaussians (called
> with rama) and quartics (also called by rama).
Just to fill in the background for others: The initial
implementation of the RAMA term would
read in a grid of potential values at a fixed (usually 2-5 degrees)
resolution. But it had fairly
rough atomic forces, and used a lot of memory. So I changed the
implementation to read
in fitted curves (either using Gaussian or quartic functions)
generated from the raw potential
grids. These are much less memory intensive and much smoother (and
therefore easier to
optimize).
> When I call rama from
> python, what set am I using: gaussians or quartic? And am I using
> short
> range (intra-residue) as well as long range (inter-residues)
> correlations?
Look at the code in xplor/python/protocol.py, starting around line
570. It uses the 2D and 3D
potentials corresponding to intraresidue correlations only, fitted
with quartic functions.
> One more. Reading the article [J. Magn. Res. 146, 249-254 (2000)],
> the
> authors define the DELPHIC torsions database as the original
> implementation of the torsion angle potential. So, when talking about
> the DELPHIC database, are we talking about the raw potentials or
> something else?
I intended the term 'DELPHIC torsions' to apply to all
implementations of the method,
whether they use the raw grids, or fitted curves. The name 'RAMA'
came about because
it was initially just for backbone phi/psi angles. Unfortunately, it
became widely used
before I thought of the much snappier 'DELPHIC' name. Sigh.
> And if that's the case, then instead of calling rama
> should I call xrama instead of rama?
If you want to use the implementation that uses raw potential grids,
you need to call the
xrama term. In python, you could just paste in the setup code from a
classic xplor script,
and wrap it in an xplor.command() call.
>
> protocol.initRamaDatabase()
> potList.append( XplorPot('XRAMA') )
Note that the above code won't work. The call to initRamaDatabase()
reads in the 2D and
3D quartic potentials and calls their corresponding setup script. It
doesn't read in the raw
potential grids, or call their corresponding setup scripts.
Can I ask why you want to use the XRAMA implementation, anyway?
Hope this helps.
--JK
