Re: Ellipsoid minimization does not converge -- May 06, 2010 - 15:29

Hi Daniel,

I agree with Ed that cycling through 2-3-4-5 different, but very similar 
solutions is an issue with this protocol. You can manually check in the 
logs with the chi2 value... The cycling should be quite evident as in 
this case we encountered recently for the sphere diffusion tensor:

i 23706.631648749521
1 462.59288287282465
2 466.46983971466807
3 466.46983971466801
4 466.4698397146683
5 466.46983983218593
6 466.46983971466784
7 466.46983971466773 cycle_I -> pseudo-CONVERGENCE sphere
8 466.46983971466801 as_3 cycle_I
9 466.46983982041428 cycle_I
10 466.46983971466864 cycle_I
11 466.46983971466778 cycle_I
12 466.46983971466773 as_7 cycle_II
13 466.46983971466801 as_3_8 cycle_II
14 466.46983982041428 as_9 cycle_II
15 466.46983971466864 as_10 cycle_II
16 466.46983971466778 as_11 cycle_II
17 466.46983971466773 as_7_12 cycle_III
18 466.46983971466801 as_3_8_13 cycle_III

In this example, we chose the first repetitive part as the converged 
iteration (i.e. the first run in the first cycle).

...

If this can be of any help, I suggest you have a look at the 
supplementary material of a paper we recently published in Biophysical 
Journal:

Fisette, Morin, Savard, Lagüe & Gagné (2010) TEM-1 Backbone Dynamics — 
Insights from Combined Molecular Dynamics and Nuclear Magnetic 
Resonance. Biophysical Journal, 98: 637–645.

Table S3 on page 21 shows the details of different runs of this protocol 
with different slightly different protein structures... As you will see, 
the number of iterations for the ellipsoid model was between 4 and 12 
for all runs, except for a run not converged after >30 iterations... 
hence, the number of iterations is very variable...

Good luck !


Séb :)




On 10-05-05 5:03 AM, Edward d'Auvergne wrote:
> Hi,
>
> For the ellipsoid model, my experience is that ~15 rounds of
> optimisation are required.  I can't reinforce enough how complex the
> model-free combined optimisation and model selection problem actually
> is (all explained in d'Auvergne E. J., Gooley P. R. (2007). Set theory
> formulation of the model-free problem and the diffusion seeded
> model-free paradigm. Mol. Biosyst., 3(7), 483-494.
> http://dx.doi.org/10.1039/b702202f)!  This can take up to 2 weeks
> depending on how fast your computer is (at 2.4 GHz, I would expect 2
> weeks).  Note that later iterations are much faster than the start.  I
> haven't tested the convergence of the global model with simplex
> optimisation, so I have no idea how long this will take and if
> convergence can always be reached.  One issue that can occur is a 3 to
> 4 iteration cycle around the universal solution (combined optimisation
> minimum and model selection minimum, see the ref).  This you can
> manually check in the results files to see if any 2 are identical.  If
> so, you can stop the calculation.  Any of the positions in the cycle
> should be ok, as the models and values should be very similar for all
> members.  I really should add this check to the script!  I hope this
> helps,
>
> Regards,
>
> Edward
>
>
>
> 2010/5/5 Daniel Pérez<daniel.perez@xxxxxxxxxxxxxxxx>:
>    
> > Hi everyone,
> >
> > I've been running a full_analysis script with relaxation data at two fields 
> > (600 and 900MHz). The calculations with models I to IV run successfully. 
> > However, the ellipsoid model does not converge even after 7 rounds and more 
> > than one week of calculation time.
> >
> > I checked the mailing list and found one similar situation that happened to Sebastien (October 
> > 24, 2007. Vol.17 issue 8 ... Extremely long optimization times). I followed the suggestion by 
> > Edward in order to change the minimization protocol from "newton" to 
> > "simplex". The result is that after two weeks and 27 rounds the model ellipsoid still 
> > does not converge, and with each next round the minimization becomes slower and slower.
> >
> > The protein is an homodimer of 36KDa. For the model free analysis I'm using 
> > the pdb structure of the monomer (18KDa), and I might have wrong assigned 
> > resonances.
> >
> > I'm running relax 1.3.4 on 2 cores Intel(R) Core(TM)2 CPU,  6600 @ 2.4GHz, 4M 
> > cache, and 4G ram.
> >
> > Do you have any idea?. I'll appreciate any suggestion to follow with the 
> > analysis.
> >
> > Thanks
> >
> > Daniel
> >
> > Dr. Daniel Pérez
> > Institut für Molekularbiologie und Biophysik
> > Schafmattstrasse 20
> > ETH-Hönggerberg, HPK G 5
> > CH-8093 Zürich
> > Phone: +41 44 633 25 37
> > E-mail: daniel.perez@xxxxxxxxxxxxxxxx
> > _______________________________________________
> > 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
> >
> >      
> _______________________________________________
> 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
>    

--
Sébastien Morin
PhD Student
S. Gagné NMR Laboratory
Université Laval&  PROTEO
Québec, Canada