I certainly expect Chi-squared and the dynamic parameters to be highly
dependent on the value of tm. This is why I recomend the extensive
iteration of steps 2-4 below, in an attempt to find the optimal fit. I
would expect if you start the procedure below from subtly different tm
values and iterate to convergence you should reach the same result in
the end (ie. tm, chi2 and all other parameters should be nearly
identical). If that is not the case, you could use the model-selection
functions in relax to test which result is the best fit to your data.
This is a little dangerous, however, because you have no way of knowing
that the two possible solutions you are considering are the only
possibilities (ie. there might always be another, better, solution that
you havn't found yet).
One alternative is to use the analysis protocol implimented in the
full_analysis.py sample script. This is a new and quite different
approach, that does not rely on having an initial tm estimate. It has
been discussed in the thread Edward pointed you to, and elsewhere on
this list.
One final point to keep in mind is that all modelfree analysis protocols
can be effected by bad data. It is well worth looking carefully for
apparent outliers, residues that appear to be strongly affected by Rex,
etc. and excluding them from the early stages of the analysis. They can
always be reintroduced after you have settled on a final diffusion
tensor.
Chris
On Mon, 2007-01-08 at 10:04 +0800, Hongyan Li wrote:
Dear Chris,
Thanks for the helpful suggestion.
I have tried as you suggested to repeat steps 2-4 from estimated tm and then
from best-fit tm. Since estimated tm I used is from modelfree (which is very
good) I actually got converged results immediately. However, I noticed that
a
subtle difference in tm caused Chi-square significantly different. Of cause,
other parameters are also different. The question is how to judge which set
of
data is more accurate (based on Chi-square??).
Best wishes,
Hongyan
Quoting Chris MacRaild <c.a.macraild@xxxxxxxxxxx>:
Hi Hongyan,
relax is designed to be completely flexible in the way you perform your
analysis, allowing for the procedure to be tailored to the system at
hand, or for new proceedures to be developed. One procedure that I can
recomend which is somewhat similar to the one you outline is as follows:
1. estimate tm
2. fit each residue to dynamic models
3. select best model
4. fit tm and dynamic parameters simultaneously
5. repeat steps 2-4 starting from best-fit tm value. Continue until
results converge
6. repeat steps 2-5 for each diffusion model (isotropic, axially
symetric and anisotropic)
7. select best diffusion model
8. Monte Carlo simulations (error analysis)
As you note, Monte Carlo simulations over all parameters will be very
slow. This is why I recommend only performing the error analysis at the
end of the whole proceedure. I some cases it may be necessary to perform
the Monte Carlo simulations over only the dynamic parameters (ie. with
diffusion tensor fixed) in order to improve efficiency.
There has been some discussion of this and other analysis proceedures on
this list before. The thread that starts here:
https://mail.gna.org/public/relax-users/2006-10/msg00007.html
is worth a look.
Chris
Dr. Hongyan Li
Department of Chemistry
The University of Hong Kong
Pokfulam Road
Hong Kong
Re: scripts