Re: Stochastic velocity rescaling for Free Energy calculations

From: Brian Radak (brian.radak_at_gmail.com)
Date: Mon Feb 25 2019 - 12:05:22 CST

Hi Braden,

The main virtue of stochastic rescaling is that it somewhat minimally
perturbs dynamics (especially compared to Langevin dynamics) - I refer you
to Michael Shirts' paper in JCTC a few years back showing that Newtonian
dynamics and stochastic rescaling (as commonly used) yield essentially
identical diffusion constants.

In the CUDA version of NAMD the vast reduction in PRNG and SHAKE calls also
yields a significant speedup, although future developments should decrease
this advantage.

There is only one special consideration for alchemical free energy
calculations, which is the behavior of the alchemical region at the
endpoints. Since stochastic rescaling does not perturb the direction of the
velocities (only their magnitude), non-interacting particles will
essentially diffuse in a straight line in a semi-deterministic fashion -
this can be horrible for sampling. Langevin dynamics, on the other hand,
will shake the particles around at random. Really, this is just a
consequence of the nonsense of MD of ideal gas particles, but the lazy
practitioner usually just relies on that randomness to be efficient.
Replica exchange may mitigate the issue, as well as additional dummy
interactions or restraints (these are usually necessary anyway). You could
also be really fancy and dump MD for a minimal Monte Carlo placement of the
alchemical atoms.

HTH,
BKR

On Sat, Feb 23, 2019, 4:23 PM Braden Kelly <bkelly08_at_uoguelph.ca> wrote:

> Hello,
>
> I just saw that a new feature in namd is the stochastic velocity
> rescaling (thermostat (Bussi et al.).
>
> Has anyone tried it for alchemical free energy calculations? (I am looking
> at you Brian since I know you do those 🙂 )
>
> Common knowledge tends to be that Langevin is the best thermostat for Free
> Energy calculations. Does this new algorithm compare favorably with
> Langevin dynamics free energy calculations?
>
> Thanks,
>
> Braden
>
> *Braden Kelly*
>
> PhD. Candidate, E.I.T
>
> University of Guelph
>
> Biophysics Interdisciplinary Group (BIG)
>
> "I feel more like I do now than I did when I first got here."
>
>
> [image: 1495766834308_mini_wolf_noir.jpg]
>


Outlook-1495766834.jpg

This archive was generated by hypermail 2.1.6 : Sat Dec 07 2019 - 23:20:29 CST