Hi,
The first part of your post is simple to respond to as I have used
exactly this approach in my PhD thesis. This is described in Chapter
7 of the thesis which can be downloaded from
http://eprints.infodiv.unimelb.edu.au/archive/00002799/ . The
hybridisation function docstring 'help(run.hybridise)' will explain
the rest (otherwise if you could help me in expanding this
documentation so that it is fully comprehensible, that would be of
great use (those coding are often blind to deficiencies in the code's
documentation)).
The second part is a little more worrying. The results should have
converged well before 30 iterations unless there is something
seriously wrong with the model or data. The inclusion of residues for
which there is only 3 data sets may be the problem but, from memory,
as you are using models with many more parameters than this these
should be automatically deselected by relax. There is one scenario
that I can theoretically conceive of and this has to do with the
full_analysis.py script's attempt at finding the solution within the
universal set by incorporating mathematics through the optimisation of
the chi-squared function while also optimising the statistical
quantity known as the Kullback-Leibler discrepancy. If this makes no
sense, my publications (when they are all out) or thesis should
explain everything. The scenario is that the dual optimisations are
feeding off each other and causing the results to flip-flop between
two continually interchanging models. There could simply be one
parameter, being close to insignificance, that is appearing and
disappearing causing the chi-squared and AIC values to be repeated
every two iterations. Simply tabulating the chi-squared value for a
number of these iterations should clearly demonstrate this problem.
All that being said, this scenario is quite unlikely and something
else is more likely to be the problem. Again a list of the
chi-squared values for a large number of iterations would be very
useful in tracking down the issue.
Regards,
Edward
On 6/24/07, Douglas Kojetin <douglas.kojetin@xxxxxxxxx> wrote:
Hi All,
I have two separate but related questions. I am using relax 1.2 (svn
version 3301). I have relaxation data collected at two fields -- 500
and 600 MHz. However, I have data for 9 additional residues at 600
MHz that were unresolved at 500 MHz.
(1) The protein I am studying has two domains, with considerable
interactions between them, connected by a flexible linker. When all
data (domains + linker) was included in the calculations, the
full_analysis.py protocol picked local_tm for the AIC selection of
the diffusion tensor. I would like to analyze my data using a hybrid
model: (a) the two domains together (using the same diffusion
tensor) and (b) the flexible linker using a separate diffusion tensor
(likely local_tm). My guess is that a prolate or oblate tensor will
be selected for the domains when analyzed without data from the
linker region (the quadric_diffusion program from Art Palmer suggests
an axially symmetric tensor is a good approximation).
Can anyone provide an example of a script where relax is used to
analyze a hybrid model, or briefly outline the steps? For example,
should I run a local_tm optimization using all residues, then
unselect the flexible linker residues in the unresolved file (as
specified in the full_analysis.py script) and continue the
optimization of the other tensors (sphere, prolate, oblate and
ellipsoid)?
(2) I am currently running the full_analysis.py protocol, without
the data for the linker region. The optimization of the prolate
tensor is taking much longer than the other tensors for this
calculation (currently on round_30), as well as the prolate
calculation using all data including the linker region (it converged
in 14 rounds). The differences in the parameters between rounds are
very small:
"""
#####################
# Convergence tests #
#####################
Chi-squared test:
chi2 (k-1): 785.88714033105236
chi2 (k): 785.88714033128417
The chi-squared value has not converged.
Identical model-free models test:
The model-free models have converged.
Identical parameter test:
Parameter: tm
Value (k-1): 6.794068350295769e-09
Value (k): 6.7940683502957698e-09
The diffusion parameters have not converged.
Parameter: Da
Value (k-1): 6337661.7164024841
Value (k): 6337661.7164041474
The diffusion parameters have not converged.
Parameter: theta
Value (k-1): 1.6904048161417038
Value (k): 1.6904048161417222
The diffusion parameters have not converged.
Parameter: phi
Value (k-1): 0.30710640562938446
Value (k): 0.30710640562950142
The diffusion parameters have not converged.
"""
https://mail.gna.org/public/relax-devel/2007-06/msg00012.html
relax does not report a problem for a specific residue, as was
reported in the following post (https://mail.gna.org/public/relax-
users/2006-12/msg00002.html). Could this be a result of having data
at only one field for the 9 residues?
Thanks in advance,
Doug
_______________________________________________
relax (http://nmr-relax.com)
This is the relax-users mailing list
relax-users@xxxxxxx
To unsubscribe from this list, get a password
reminder, or change your subscription options,
visit the list information page at
https://mail.gna.org/listinfo/relax-users
Re: two questions: hybrid analysis and convergence problems