- Contemporary messages sorted: [ by date ] [ by thread ] [ by subject ] [ by author ] [ by messages with attachments ]

From: Jason Swails <jason.swails.gmail.com>

Date: Tue, 20 Oct 2015 10:11:01 -0400

On Mon, Oct 19, 2015 at 9:22 PM, B. Lachele Foley <lfoley.ccrc.uga.edu>

wrote:

*> "I'm sure she'd be happy to share her slides if she can find them and you
*

*> wanted them. </throwing Lachele under the bus>"
*

*>
*

*> Not that I mean to avoid being tossed under a bus, and certainly neither
*

*> do I wish to cause anyone else such discomfort, but... I just looked at my
*

*> slides. Are you sure you meant me? From 2015?
*

*>
*

I am 100% sure that I meant you when I wrote that. :) But I think I'm

remembering differently since you clearly would know better than I whether

you gave that presentation -- I actually think it was Dave Cerutti that

gave that presentation now. But I *know* it was given (and given by

someone working on FFs, obviously), because I remember asking a question

about it. I also remember a slide where blindly fitting a Fourier series

without the n=0 term gave bad results and including the n=0 term gave a

better fit, precisely because there was an energy shift difference between

the MM and QM profiles. Another way of working around this, as Carlos

said, is to simply shift the QM profiles down to match where the MM profile

is scanning (basically, make the minimum energy of the QM scan 0, I

think). This seems easier to me than trying to center the QM scan around

the origin so you can fit it to cos(n*phi) instead of (1+cos(n*phi)).

Dave -- do you still monitor this list? Can you verify that I'm not losing

my mind? :)

But, without having time to think as deeply as I would like about it (and

*> hoping Ross will... :-), your suggestion seems ok. I suppose it could add
*

*> extra steps to the calculation, but perhaps a couple lines of code to
*

*> reduce the terms could fix that, too.
*

*>
*

It shouldn't really add any steps to the calculation -- it's just adding

another term in the curve fit. But Adrian and Chad Hopkins worked out a

way to fit dihedrals using a linear least squares fit -- none of this

nonlinear stuff that can get stuck in a fake local minimum; something that

gives you the exact result incredibly quickly. I think they've published

this work as well. But if you just fit all terms from n=0 to n=6 (instead

of n=1 to n=6), your fit builds in an arbitrary potential shift.

I thought it was there to avoid having a negative correction in the torsion

*> terms. This is what I told some students the other day... But, I'm not
*

*> convinced that will necessarily be a problem. Re-tooling all the force
*

*> fields, however, could require some work.
*

*>
*

All that removing the 1+ would do is introduce a constant shift in the

potential, which doesn't matter when it comes to computing forces and

thermodynamic quantities. But it *would* make Amber the black sheep of the

MD codes, which is a very good reason not to do it (of course it doesn't

matter how you *fit* the terms, so long as you're fitting a function with

the same gradients as the Amber potential everywhere).

All the best,

Jason

Date: Tue, 20 Oct 2015 10:11:01 -0400

On Mon, Oct 19, 2015 at 9:22 PM, B. Lachele Foley <lfoley.ccrc.uga.edu>

wrote:

I am 100% sure that I meant you when I wrote that. :) But I think I'm

remembering differently since you clearly would know better than I whether

you gave that presentation -- I actually think it was Dave Cerutti that

gave that presentation now. But I *know* it was given (and given by

someone working on FFs, obviously), because I remember asking a question

about it. I also remember a slide where blindly fitting a Fourier series

without the n=0 term gave bad results and including the n=0 term gave a

better fit, precisely because there was an energy shift difference between

the MM and QM profiles. Another way of working around this, as Carlos

said, is to simply shift the QM profiles down to match where the MM profile

is scanning (basically, make the minimum energy of the QM scan 0, I

think). This seems easier to me than trying to center the QM scan around

the origin so you can fit it to cos(n*phi) instead of (1+cos(n*phi)).

Dave -- do you still monitor this list? Can you verify that I'm not losing

my mind? :)

But, without having time to think as deeply as I would like about it (and

It shouldn't really add any steps to the calculation -- it's just adding

another term in the curve fit. But Adrian and Chad Hopkins worked out a

way to fit dihedrals using a linear least squares fit -- none of this

nonlinear stuff that can get stuck in a fake local minimum; something that

gives you the exact result incredibly quickly. I think they've published

this work as well. But if you just fit all terms from n=0 to n=6 (instead

of n=1 to n=6), your fit builds in an arbitrary potential shift.

I thought it was there to avoid having a negative correction in the torsion

All that removing the 1+ would do is introduce a constant shift in the

potential, which doesn't matter when it comes to computing forces and

thermodynamic quantities. But it *would* make Amber the black sheep of the

MD codes, which is a very good reason not to do it (of course it doesn't

matter how you *fit* the terms, so long as you're fitting a function with

the same gradients as the Amber potential everywhere).

All the best,

Jason

-- Jason M. Swails BioMaPS, Rutgers University Postdoctoral Researcher _______________________________________________ AMBER-Developers mailing list AMBER-Developers.ambermd.org http://lists.ambermd.org/mailman/listinfo/amber-developersReceived on Tue Oct 20 2015 - 07:30:05 PDT

Custom Search