Re: AW: AW: NAMD feature request Hydrogenscale

From: Vincent Leroux (vincent.leroux_at_loria.fr)
Date: Thu May 31 2012 - 10:40:32 CDT

Hi Norman,

Of course a timestep of 4 fs for electrostatics will harm energy conservation, but usually not critically so. You may try to increase cutoff distances to see if that minimizes the problem, but you may lose a good part of the speed increase in the process. But frankly, correct me if I am mistaken, there is not much you can do about that, this is bound to integrator stability and this is the price to pay for timesteps larger than 1 fs. If you are worried about a 1 K/ns temperature raise, Langevin NVT dynamics may solve the problem with just a very small damping factor (such as 0.01), so you introduce temperature control as lightly as possible. The temperature will fluctuate slightly around the target but this should be acceptable. Another workaround would be velocity rescaling with large intervals (e.g. 1 ns) but you have to be careful of the "hot solvent cold solute" issue then. At this point, you may wonder what is best between this and a neat NVE equilibrium production run with 1 fs timesteps, but 3-4x less
 accessible simulation time. Depends on what you want to see.

And frankly, I am fundamentally not convinced by hydrogen mass repartitioning. Question of philosophy, maybe, but the CHARMM and AMBER FFs were carefully parameterized against reference QM data with "normal" protons. It may look like heavier protons are an easy way to increase simulation speed while (apparently) conserving energy, but IMHO if you fall for it you move your model further away from reality to the point it is now quite unphysical, and you will pay a price for that too. In the best case scenario, I guess you will simply mess up the H-bond networks at biomolecular interfaces, especially those involving waters. But forcefield parameters are interdependent to a very large extent: change some atomic masses, and I bet you have to re-parameterize a lot of other parameters, which is downright crazy. The CHARMM or AMBER devs may better warn you against that better than me, or at least describe precisely what to expect and not to expect from this trick.

If you do not like the protons moving too much why not try an unified-atom FF? It will by nature tolerate larger timesteps. There are other "maybe good enough" solutions for faster simulations that do not involve changing physical constants. Implicit solvent is the most obvious, and some AMBER forcefields seem quite good at that. The recent Molly approach from NAMD may be appropriate, have not tried it yet. If you want the best conformational exploration possible unbiased MD may also not be the optimal strategy.

Regards
VL

----- Mail original -----
> De: "Norman Geist" <norman.geist_at_uni-greifswald.de>
> À: "Vincent Leroux" <vincent.leroux_at_loria.fr>
> Cc: "Namd Mailing List" <namd-l_at_ks.uiuc.edu>
> Envoyé: Jeudi 31 Mai 2012 08:06:29
> Objet: AW: AW: namd-l: NAMD feature request Hydrogenscale
>
> > -----Ursprüngliche Nachricht-----
> > Von: owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] Im
> > Auftrag von Vincent Leroux
> > Gesendet: Mittwoch, 30. Mai 2012 17:28
> > An: namd-l_at_ks.uiuc.edu
> > Betreff: Re: AW: namd-l: NAMD feature request Hydrogenscale
> >
> > Excuse me, the fact that a method is published does not prevent me
> > to
> > criticize it.
>
> Hopefully not. No opinion, no progress.
>
> >
> > Well, if you do not want to compute more timesteps than required,
> > then
> > why don't you use r-RESPA-based multiple timestepping (nonbonded 2
> > and
> > fullelectfrequency 4 can work very well on your typical
> > protein-protein
> > or protein-ligand system provided careful equilibration was done)
> > and
> > the Molly method as recently implemented in NAMD?
>
> I'm already using fullelectfrequency 4 for better speedup through the
> gpus, but that harms my energy conservation and my system increase
> temperature by about 0.5K-1K per ns.
>
> >
> > This looks IMHO a little bit more serious than just changing atom
> > masses
>
> Question of philosophy. I don't think calculating non-bonded and
> electrostatic forces with atom positions that are no more true is
> better.
>
> > while using a forcefield that was not parameterized against these
> > new
> > unnatural values. In the reference you provide, it is just
> > mentioned
> > such a trick only moderately affects the viscosity in a TIP3P water
> > box,
> > you have to go to the original reference from the people that
> > actually
> > had the idea to get some serious validation attempts:
> > http://dx.doi.org/10.1002/(SICI)1096-987X(199906)20:8<786::AID-
> > JCC5>3.0.CO;2-B
>
> Please read more carefully. The viscosity change of the water was a
> side effect, not the effect that was targeted.
>
> >
> > You will see that the gain in performance is not so high, without
> > resorting to additional tricks, and that the effect of the
> > dynamical
> > behavior of the simulated system is not negligible...
>
> That's not true. Without additional tricks a timestep of 4fs is
> possible only with the hydrogen mass repartitioning, 4 times higher
> than namds default 1fs.
>
> >
> > Do you still think this is a good idea? If you wish, go ahead and
> > ask
>
> Yes.
>
> > the CHARMM forcefield developers their opinion... this is
> > especially
>
> In fact I use the amber ff.
>
> > relevant since according to the Berendsen paper hydrogen mass
> > repartitioning only works well if piled up with dummy atoms usage,
>
> Not true. This only provides the possibility of even higher timesteps
> than 4fs.
>
> > requiring some additional tinkering with the force field itself, so
> > that
> > the dihedrals are not messed up... So you need a kind of forcefield
> > able
> > to tolerate that. That would certainly be more GROMOS than CHARMM.
> >
> > Regards
> > VL
> >
> >
> > On 30/05/2012 16:17, Norman Geist wrote:
> > > And of course a link to the publication of this method as it is
> > already used.
> > >
> > > Accelerating Biomolecular Dynamics in the Microsecond Time Scale
> > > http://pubs.acs.org/doi/full/10.1021/ct9000685
> > >
> > > A little more serious please.
> > > Why should I compute every timestep if I don't need it.
> > >
> > > Norman Geist.
> > >
> > >> -----Ursprüngliche Nachricht-----
> > >> Von: owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu]
> > >> Im
> > >> Auftrag von Vincent Leroux
> > >> Gesendet: Mittwoch, 30. Mai 2012 15:04
> > >> An: namd-l_at_ks.uiuc.edu
> > >> Betreff: Re: namd-l: NAMD feature request Hydrogenscale
> > >>
> > >> Focusing on the potential "scientific value" of making all
> > >> hydrogens
> > 4x
> > >> heavier for the sake of getting MD "just run faster with big
> > >> timesteps",
> > >> how about passing the idea to the CHARMM forcefield developers
> > >> in
> > the
> > >> dedicated mailing list? ;-)
> > >>
> > >>
> > >>
> > >> On 30/05/2012 14:14, Axel Kohlmeyer wrote:
> > >>> On Wed, May 30, 2012 at 7:16 AM, Ajasja Ljubetič
> > >>> <ajasja.ljubetic_at_gmail.com> wrote:
> > >>>>> without commenting on scientific value,
> > >>>>
> > >>>>
> > >>>> Could we hear this comment as well? (If it is not to vulgar,
> > >>>> of
> > >> course:)
> > >>>
> > >>> since i have no proper justification for
> > >>> it, i'd rather not. the world is already
> > >>> full of too many prejudices and people
> > >>> with half-knowledge or no knowledge
> > >>> blasting out opinions and speculations.
> > >>>
> > >>> axel.
> > >>>
> > >>>>
> > >>>> Best regards,
> > >>>> Ajasja
> > >>>
> > >>>
> > >>>
> > >
> > >
>
>

This archive was generated by hypermail 2.1.6 : Tue Dec 31 2013 - 23:22:02 CST