2009/12/14 Tiago Pais <tpais@xxxxxxxxxxx>:
I am not really sure I am following you.
Let's see and please correct me if and where I may be mistaken. To optimize
de diffusion tensor model I fix the modelfree models (m0-m9) and let the
diffusion tensor be optimized. I have to questions with regard to this:
1- Does this optimization tells you if it is isotropic, axially symmetric or
fully anisotropic? Where?
This will be what you set it to earlier with the
diffusion_tensor.init() user function. Optimisation doesn't tell you
this, but from the number of parameters minimised you can tell:
1 - sphere
4 - spheroid
6 - ellipsoid
2- Is there an exclusion of spins for tensor optimizations during this
procedure, i.e., are all spins used to optimize the diffusion tensor? Or
should I run this optimization using only a subset of residues that comply
with the criterion of reduced mobility?
All selected spins are used. All deselected spins are excluded.
However prior to optimisation, spins with no data, not enough data,
etc. are deselected (now you will get a RelaxWarning explaining the
problem).
For optimising a diffusion tensor during a model-free analysis, all
possible spins should be used. The exclusion of spins is only for
getting the diffusion tensor from the R1/R2 ratio (from Kay et al.,
1989). However this R1/R2 approach for proteins is problematic. E.g.
in cases where the NH vectors of certain secondary structure elements
point along the long axis of the protein, the high R2s might be
mistaken as mobility and excluded. Then the anisotropy of the tensor
is underestimated and the repetitive optimisation may not be able to
slide to the correct solution. This catastrophic failure is what I
showed with the Olfactory Marker Protein (OMP) in my PhD thesis
(downloadable as PDF at
http://www.nmr-relax.com/refs.html#dAuvergne06, and also
http://www.amazon.com/Protein-Dynamics-Model-free-Analysis-Relaxation/dp/3639057627/ref=sr_1_1?ie=UTF8&s=books&qid=1260815221&sr=8-1)
and also published in my second relax paper
(http://www.nmr-relax.com/refs.html#dAuvergneGooley08b).
The R1/R2 ratio also experiences catastrophic failure when there is a
large amount of mobility in the system. This was demonstrated by
Orehkov et al, 1999 with the Bacteriorhodopsin fragment (1-36)BR.
This is again in my thesis in section 6.4.4, and
http://www.nmr-relax.com/refs.html#dAuvergneGooley08a.
This is the reason I developed the new protocol used in the
full_analysis.py script. Specifically to avoid these huge analysis
failures, but also to remove the mess that is the initial diffusion
tensor estimate. In model-free analysis there is a chicken and egg
question. Do you start the with the diffusion tensor or start with
the internal motion? Starting with both is not physically possible
due to the fact that you are searching for the best solution across
4*n^10 models, where 4 are the 4 diffusion tensor types, n is the
number of spins analysed, and 10 is the number of model-free models
used. This is a massive number of spaces to search through, and the
number of dimensions in each individual optimisation space can also be
huge! So you have a problem whereby you are searching the solution
not only within one optimisation space, but across incredible numbers
of optimisation spaces. It is a joint optimisation and
model-selection problem. I've described this all in far more detail
in my Mol. Biosyst. paper
(http://www.nmr-relax.com/refs.html#dAuvergneGooley07).
With this new protocol, I have reversed the concept. Rather than
starting with the diffusion tensor, here you are starting with the
internal motions. This however required multiple field strength data,
but to avoid the catastrophic failures, you also have to have multiple
field strength data. My 2007 and two 2008 papers (or the thesis)
explains this much better and how this protocol in the
full_analysis.py script tries to find the universal solution (the one
solution in the universal set of all 4*n^10 spaces) in a more rigorous
way than is currently done in the NMR field.
Sorry for all the trouble before you leave for holydays.
Not a problem. I hope this long email helped.
Regards,
Edward
Re: AIC to select diffusion model