On Feb 17, 2008 7:47 PM, Sebastien Morin <sebastien.morin.1@xxxxxxxxx> wrote:
Hi,
I recently used the full_analysis.py script.
In the final run, I wonder if it would be useful to make MC simulations
on the diffusion tensor parameters. Are errors on the diffusion tensor
parameters useful ?
They are, but it takes a long time. relax's more accurate and
significantly higher precision optimisation, compared to Modelfree and
Dasha, means this takes a long time. Especially when this is combined
with the phenomenon of model and MC sim failure (the paper on
Model-free Model Elimination is published in JBNMR). You can give it
a go, but it could take forever - literally. There is an ancient post
on this, lets see if I can dig it out. Ok, this is from Chris
MacRaild, way back in 2005. Well before the mailing lists. So I
have cut and paste the relevant part of my response, it's located at
the end of this message. My model elimination paper has a very good
description of this problem too.
I ask this question because I used to work with ModelFree (Palmer) and
this program output errors on every parameter, including the diffusion
tensor...
Also, I realized that simulations were being done on the bond length and
CSA. Is this useful ? Why ?
If the bond length or CSA parameters are part of the model, which is
not the case the full_analysis.py script or if you are doing a normal
analysis, then these will be treated as all model parameters and will
be part of the simulation process (just as it was in the fitting).
This is a hidden feature of relax. Someone adventurous could be the
first in the field to play with these model-free 'constants' converted
to variables, but it will require heavy testing. Synthetic test
models are an absolute must! Maybe one day in a future post-doc or
beyond, I'll get around to this.
What's your opinion about MC simulations on the diffusion tensor
parameters as long as the bond length and CSA in the final run ?
MC simulations of the diffusion tensor will require Gary Thompson's
multi processor code for running this on a cluster. Otherwise it will
likely be impossibly slow (unless you crank down the accuracy and
precision to that of Modelfree or Dasha (I wouldn't recommend this as
it will be far worse than not simulating the tensor parameters)). The
bond length and CSA values are constant in the models you are using,
so they are unaffected by the simulations.
Regards,
Edward
Here is part of the ancient message:
-----
On 08 Sep 2005 17:37:09 +0000, Chris MacRaild wrote:
On Thu, 2005-09-08 at 08:36, Edward d'Auvergne wrote:
Hi,
That's a tough question to answer. I'm not exactly sure what you are
trying to do, but is it that you would like to constrain the longest
of the three axes of the diffusion tensor to a general position in
space where logic would suggest it should be?
Not quite. The homodimer has a two-fold axis of rotational symmetry, so
one axis of an asymmetric diffusion tensor must be co-linear with that
molecular symmetry axis. This need not be the longest axis - it could be
any of the three. This is an absolute physical constraint, and what we
are trying to do is minimise the space over which we search for the
diffusion tensor based on this constraint.
If so, does
minimisation of the prolate or fully anisotropic tensors not return
reasonable results? Or do they never converge?
Based on T1/T2 ratios, we are fairly sure that a fully anisotropic
tensor is required, and the tensor that comes out of that analysis is
consistent with what we expect (ie. one axis lies along the molecular
symmetry axis). An unconstrained analysis of the fully anisotropic
tensor is simply too slow in relax, so I don't know if relax would also
converge on a reasonable result without constaints. Adding the
constraint will make the analysis much faster, so thats a major reason
why we are so keen to impliment it.
I think the problem is in the Monte Carlo simulations. The way it is
set up in the script is that for each simulation, the global model
(diffusion tensor with all model-free parameters for all residues
simultaneously) is optimised. This is probably not the best idea
though. Although you won't get diffusion tensor errors, if you
minimised each residue separately with the diffusion tensor fixed, you
will probably get better results (and it will be orders of magnitude
faster). The reason is because some simulations fail and need to be
removed. Remember how I told you that certain models fail which is
why the script implements the eliminate function before model select,
well simulations also fail. The current version of relax has the
elimination of simulations implemented. I'm about to send off my
paper on model elimination which talks in depth about this, if you are
interested in reading a pre-press copy. The only downside of having
the diffusion tensor fixed is the bias introduced - although every
single analysis in existence today has this bias so no one will
notice. The result is a slight underestimation of the errors -
although I don't know the techniques required to fix this problem.
If computation time is the only issue, just add the appropriate
commands to the script to fix the diffusion tensor prior to the Monte
Carlo simulations, and all will be okay.
Re: full_analysis.py