Hi,
2009/10/5 Boaz Shapira <bshapira@xxxxxxxxxxxx>:
The program C++OPPS 1.0.0 is for 15N NMR relaxation.
Ah, ok. If the target function is known (the thing optimisation uses
whereby the model parameters are pumped in and the function returns a
single value - usually the chi-squared), then this functionality could
easily be added to relax. The absence of this crucial target function
in all their papers prior to 2005 is the reason I never added SRLS to
relax. If known, anyone could add SRLS into relax (note I don't have
the resources or time for this). The key here is a set of equations
converting the SRLS parameters into the 5 spectral density values
J(0), J(wN), J(wH+wN), J(wH), J(wH-wN)! The first and second partial
derivatives will then be needed (gradient and Hessian respectively).
Comparison at this point would means analyzing by both softwares that claim
to fit experimental data to a particular model, and see if there is any
consistence.
Consistence can be something like:
1) Agreement between the suggested models in both softwares (SRLS, as you
know, includes the model free).
It has been said that SRLS in the decoupling limit is the same as
model-free analysis, and for spherical Brownian motion of the particle
I would probably agree. However, how the diffusion tensor is treated
is different in both theories so I'm not quite sure if this statement
is entirely correct (the diffusion tensor is the most important
component of relaxation). Only a full and complete analysis using
both theories, and a conversion of the total chi2 value for the entire
system into, say, the model comparison Akaike's Information Criteria:
AIC = chi2 + 2k,
where k is the total number of parameters in the total model (sum of
the models of all spin systems), can the 2 results be directly
compared for agreement.
2) Correlation between poor fitting with relax (model free) and high
fitting with C++OPPS that suggest values outside the scope of model-free.
Until someone performs a thorough comparison using relax (no one has
yet), to assume that model-free analysis will loose could be dangerous
;)
Obviously if neither 1) and/or 2) will agree it won't say much - because one
could blame the fitting approach that is different.
relax can almost always find the universal solution (a simultaneous
optimisation and model selection problem) for model-free analysis when
using the 'full_analysis.py' script. I.e. it searches for the minimum
of the best model - essentially minimising AIC. Therefore if the AIC
value from SRLS is lower, then that is a better model for the motions
of that molecule. Vice versa if the model-free AIC value is lower,
this could be either that model-free theory is a better description or
that the SRLS data analysis tools are not advanced enough yet.
Note that after fixing all the problems with the model-free data
analysis, it now takes 1 to 2 weeks of non-stop calculation to
complete. SRLS doesn't take this long, but maybe it should! This is
where a model comparison using the model comparison tools called AIC,
BIC, AICc, ICOMP, CV, Bootstrapping, etc. can be used.
On the other hand if there is an agreement, I think it is an indication that
true physics is involved and perhaps after all we are not in the monkey
business.
If the AIC (or ICOMP, or BIC, etc.) values are similar then yes. But
then you would have to carefully check to see if the dynamic
description in both are similar.
Also here I agree that the real comparison would take place when SRLS and
model frees are reside within the same fitting package.
This will require someone taking the time to find the elusive SRLS
target function (the SRLS parameter to J(w) formulae), and adding this
to relax. This will help as then relax's advanced data analysis tools
can be applied to SRLS theory as well to obtain a better SRLS dynamic
description.
Regards,
Edward
Regards,
Boaz
________________________________
From: Edward d'Auvergne [mailto:edward@xxxxxxxxxxxxx]
To: Boaz Shapira [mailto:bshapira@xxxxxxxxxxxx]
Cc: relax-users@xxxxxxx
Sent: Mon, 05 Oct 2009 10:55:04 -0400
Subject: Re: SRLS
Hi,
No, I have not had a chance to try this software. Which specific
package did you mean though? None of the softwares E-SpiReS, C++OPPS,
or DITE from their basic descriptions are for the analysis of standard
NMR relaxation data using SRLS theory. Maybe I would need to do some
more research into what the program is called.
For comparing relax to their SRLS program, what exactly do you mean?
Would you like a measure of which theory is better, model-free vs.
SRLS. Or which fits better, which is theoretically better, or how one
approximates the other? relax does not have SRLS implemented yet -
the foundations for adding this analysis type has been laid for anyone
interested in adding this - so comparing how good the implementation
of SRLS theory is between different programs to relax is not yet
possible.
Regards,
Edward
2009/10/5 Boaz Shapira <bshapira@xxxxxxxxxxxx>:
Hi Edward.
Polimeno has a new version of the SRLS available for download.
http://www.chimica.unipd.it/licc/
Did you tried to compare?
Boaz
Re: SRLS