Re: Dispersion Back Calculation

[Jeremy Anderson]

That is extremely useful information, thanks!

Jeremy Anderson



Ph.D. Candidate

Johns Hopkins University
Program in Molecular Biophysics
Laboratory of Dr. Vincent J. Hilser, Ph.D.
3400 N Charles St, 104 Mudd Hall
Baltimore, MD 21218

(Lab) *410-516-6757*
(Cell) 715-613-0274


On Tue, Nov 15, 2016 at 8:49 AM, Edward d'Auvergne 
wrote:

> On 15 November 2016 at 14:33, Jeremy Anderson  wrote:
> > Hi Edward,
> >
> > Thanks for the follow up.  I totally understand the reasons for not
> having
> > fixed values within a relax analysis, it seems like a special case
> relative
> > to what I've seen in the literature and increases ones ability to skew
> the
> > results to their liking, I'm doing my best to safeguard against that
> myself.
>
> No problems.  You do have to be quite careful, as it is far too easy
> to fall into an alternative reality that can nicely explain some of
> the biology.  Mapping the optimisation space is an essential tool when
> constraining certain parameters to see if you are at a well defined
> minimum.  For example it can show you if you've just chopped across a
> valley in the space and the optimiser is landing at the bottom of that
> valley where it has been cut.  If you have access to the ancient, yet
> very powerful OpenDX software, you can use the dx.map and dx.execute
> relax user functions for this.
>
>
> > I was able to implement such an analysis in python using the RD models
> from
> > relax as well as the scipy and lmfit packages to both hold the dw
> parameters
> > constant while performing the usual grid search then nonlinear least
> squares
> > minimization and perform a cluster analysis holding the rate constant
> and/or
> > the major population constant amongst all residues.  The code is a bit
> of a
> > mess at the moment but I'm hoping to clean it up and make a repository on
> > github, so I can better document what I did and so other folks can check
> it
> > out if they want.
>
> Note that I originally looked at the scipy optimisation packages for
> relax.  However I found fatal bugs in all three of the algorithms
> implemented at the time (Levenberg-Marquardt being one of them).  The
> algorithms appeared to minimise the results, but they were nothing
> like what Art Palmer's Modelfree4 found.  I was comparing relax to
> Modelfree4 for debugging at the time when implementing the model-free
> analysis component.  I don't know if anything has changed since then,
> but you really need to be wary and double check whenever you use any
> part of scipy.  Anyway, because scipy's optimisation was so terrible,
> I decided to write the minfx optimisation library
> (https://gna.org/projects/minfx/).  You'll see a record of all of this
> in my publication history ;)
>
>
> > Thanks for your assistance.
>
> You're welcome!
>
> Regards,
>
> Edward
>
___
relax (http://www.nmr-relax.com)

This is the relax-users mailing list
relax-users@gna.org

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: Dispersion Back Calculation

[Edward d'Auvergne]

On 15 November 2016 at 14:33, Jeremy Anderson  wrote:
> Hi Edward,
>
> Thanks for the follow up.  I totally understand the reasons for not having
> fixed values within a relax analysis, it seems like a special case relative
> to what I've seen in the literature and increases ones ability to skew the
> results to their liking, I'm doing my best to safeguard against that myself.

No problems.  You do have to be quite careful, as it is far too easy
to fall into an alternative reality that can nicely explain some of
the biology.  Mapping the optimisation space is an essential tool when
constraining certain parameters to see if you are at a well defined
minimum.  For example it can show you if you've just chopped across a
valley in the space and the optimiser is landing at the bottom of that
valley where it has been cut.  If you have access to the ancient, yet
very powerful OpenDX software, you can use the dx.map and dx.execute
relax user functions for this.


> I was able to implement such an analysis in python using the RD models from
> relax as well as the scipy and lmfit packages to both hold the dw parameters
> constant while performing the usual grid search then nonlinear least squares
> minimization and perform a cluster analysis holding the rate constant and/or
> the major population constant amongst all residues.  The code is a bit of a
> mess at the moment but I'm hoping to clean it up and make a repository on
> github, so I can better document what I did and so other folks can check it
> out if they want.

Note that I originally looked at the scipy optimisation packages for
relax.  However I found fatal bugs in all three of the algorithms
implemented at the time (Levenberg-Marquardt being one of them).  The
algorithms appeared to minimise the results, but they were nothing
like what Art Palmer's Modelfree4 found.  I was comparing relax to
Modelfree4 for debugging at the time when implementing the model-free
analysis component.  I don't know if anything has changed since then,
but you really need to be wary and double check whenever you use any
part of scipy.  Anyway, because scipy's optimisation was so terrible,
I decided to write the minfx optimisation library
(https://gna.org/projects/minfx/).  You'll see a record of all of this
in my publication history ;)


> Thanks for your assistance.

You're welcome!

Regards,

Edward

___
relax (http://www.nmr-relax.com)

This is the relax-users mailing list
relax-users@gna.org

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: Dispersion Back Calculation

[Jeremy Anderson]

Hi Edward,

Thanks for the follow up.  I totally understand the reasons for not having
fixed values within a relax analysis, it seems like a special case relative
to what I've seen in the literature and increases ones ability to skew the
results to their liking, I'm doing my best to safeguard against that
myself.

I was able to implement such an analysis in python using the RD models from
relax as well as the scipy and lmfit packages to both hold the dw
parameters constant while performing the usual grid search then nonlinear
least squares minimization and perform a cluster analysis holding the rate
constant and/or the major population constant amongst all residues.  The
code is a bit of a mess at the moment but I'm hoping to clean it up and
make a repository on github, so I can better document what I did and so
other folks can check it out if they want.

Thanks for your assistance.

Jeremy Anderson



Ph.D. Candidate

Johns Hopkins University
Program in Molecular Biophysics
Laboratory of Dr. Vincent J. Hilser, Ph.D.
3400 N Charles St, 104 Mudd Hall
Baltimore, MD 21218

(Lab) *410-516-6757*
(Cell) 715-613-0274


On Fri, Nov 11, 2016 at 5:07 AM, Edward d'Auvergne 
wrote:

> On 27 October 2016 at 18:10, Jeremy Anderson  wrote:
> > Hi Edward and Troels,
> >
> > Thanks for pointing me in the right direction.  I had dug around a bit in
> > the test_suite directory but wanted to make sure I was looking in the
> right
> > place before I descended into the rabbit hole.
> >
> > I got the back calculation to work using the
> > ./test_suite/shared_data/dispersion/ns_mmq_3site_
> linear/relax_results/solution.py
> > script pretty much as-is, just changing the spin parameters to my liking,
> > calculating the curve, and outputting the values (ignoring the data and
> > residuals in the output file).
> >
> > Something I didn't mention is that the reason I've been importing the
> models
> > into ipython is so I can hold parameters constant through my own grid
> search
> > and minimization functions, which I had found somewhere in the
> documentation
> > was not possible inside relax for the minimization.  I originally thought
> > this would be easier outside of relax.
> >
> > The reason for this is because I'm in a situation where I can observe
> HSQC
> > peaks in slow exchange in one variant and skewed populations of one or
> the
> > other peaks in two other variants.  I've been working on using the
> > complementary information, in this case the observed dw and the kex from
> ZZ
> > exchange experiments, to investigate multi-state exchange in all
> variants.
> >
> > The chem. shift differences of the two skewed variants match the measured
> > nicely but the rates from CPMG are ~20 fold higher.  Therefore I wanted
> to
> > check and see if a 3-state model with some parameters held constant would
> > have infinite solutions (my assumption) or pop out something interesting
> and
> > be able to distinguish between a couple models of the conformational
> process
> > that I have in mind, which seems like a long shot.
> >
> > Sorry if thats too much information/way too open-ended but I figured I
> would
> > give some context to the greater situation I have found myself in.
> Thanks
> > again!
>
> Hi Jeremy,
>
> It is true that you cannot fix a parameter in relax and optimise the
> others.  The reason is two-fold.  Firstly the minfx library (
> https://gna.org/projects/minfx/ ) does support this functionality.
> Secondly, this functionality would be highly abused and a lot of
> rubbish results will appear in the scientific literature, with a
> detrimental effect on the reputation of the whole NMR field.
>
> Also, I didn't think it was worth the time investment compared to
> expanding relax to handle multiple data types at the same time, and
> then optimising one set of parameters for all experimental data
> simultaneously.  In your case, that would be loading the ZZ exchange
> and CPMG data at the same time, and optimising the single model.  This
> would be interesting, as the two experiment types contain both
> complementary and overlapping information content.  So saying that the
> overlapping content should only come from the ZZ experiment might
> over-constrain the CPMG experiment due to any biases or experimental
> noise from that experiment.  Are you able to set up the problem in
> this alternative way in iPython?
>
> Regards,
>
> Edward
>
___
relax (http://www.nmr-relax.com)

This is the relax-users mailing list
relax-users@gna.org

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