Re: Relaxation dispersion comparison: The LM63 model results from relax, NESSY and CPMGFit. -- May 29, 2013 - 15:54

Hi again,

Please ignore the comment about the constraints algorithm not being
implemented ;)  The relax_disp branch now uses the log barrier
algorithm 
(http://en.wikipedia.org/wiki/Barrier_function#logarithmic_barrier_function)
to constrain the parameters without gradient target functions.  This
was a very simple and quick fix.

Regards,

Edward




On 29 May 2013 12:01, Edward d'Auvergne <edward@xxxxxxxxxxxxx> wrote:
> Hi,
>
> Troels, you might be interested in the following comparison.  The
> reviving Sebastien Morin's relaxation dispersion branch in relax is
> now complete.  The relax_disp branch can now correctly optimise one of
> the dispersion models, that of Luz and Meiboom 1963 for 2-site fast
> exchange for CPMG-type experiments.  This does not implement the
> gradients or Hessians yet, and parameter constraints are still to be
> added.  But nevertheless relax can find the correct parameter values.
> This should be a good test of the dispersion code in relax as the
> addition of other models - such as that for R1rho data, more
> complicated models for CPMG-type data, and numerical integration of
> the Bloch-McConnell equations - should be trivial after that
> (requiring only 20-50 lines of new code, not counting comments and
> docstrings).
>
> Through relax user functions, I can now generate the input for CPMGFit
> and NESSY.  These are the relax_disp.cpmgfit_input and
> relax_disp.nessy_input user functions.  We can now also execute
> CPMGFit within relax using relax_disp.cpmgfit_execute.  This can be
> completed later, but the idea is to follow the concept of model-free
> user functions:
>
> dasha.create
> dasha.execute
> dasha.extract
> palmer.create
> palmer.execute
> palmer.extract
>
> These are for the Dasha and Modelfree4 programs respectively.
> Implementing the 3 user functions for creating input files, executing
> the program in-line, and extracting the results from the program's
> output will allow relax to use external programs as optimisation
> engines.  This is useful for comparing the results from different
> programs and really eliminating all bugs from the dispersion field.
>
> Back to the comparison, I have used Flemming Hansen's 500 and 800 MHz
> CPMG data from:
>
> Hansen, Vallurupalli & Kay, 2008, JPhysChemB, 112: 5898-5904.
>
> which he donated to Seb to be added to relax (located in
> test_suite/shared_data/dispersion/Hansen/).  Looking at a single
> randomly chosen residue (number 70), I see different results from the
> 3 programs:
>
> Param       relax       NESSY       CPMGFit
> R2 (500)        6.806       7.639       6.866
> R2 (800)        6.679       7.387       6.866
> phi         1.259e-13       0.259   1.226e-13
> phi (500)   31464.605   26292.031   30644.496
> phi (800)   80549.390   67307.598   78449.180
> kex          4763.249    3906.180      4.683
> tau         4.199e-05   5.120e-05       0.427
> chi2          106.393     156.446     106.471
>
> tau = 2/kex (I think that CPMGFit works with ms units).
>
> Obviously NESSY is not doing so well (likely due to using the horribly
> buggy scipy optimisation code) and relax and CPMGFit find the same
> result for this model.  The slight difference between relax and
> CPMGFit is that it looks like CPMGFit assumes the same R2 value for
> all static fields - which I think would be a little strange,
> especially if fast internal motions are also present for that spin
> system.
>
> Therefore I believe that this relax branch is in a state to accept
> code for the other models.  The backends for the
> relax_disp.cpmgfit_input and relax_disp.nessy_input could be also
> modified to handle these new models to allow for rigorous comparisons
> and debugging.  The dispersion infrastructure should no longer have
> any large changes, so feel free to start playing.
>
> Regards,
>
> Edward