Adaptive tempering implementation in NAMD differs from original paper?

From: Sachin Natesh (snatesh_at_uchicago.edu)
Date: Tue Sep 22 2015 - 19:45:22 CDT

Hello all,

This post concerns the NAMD implementation of the adaptive tempering
algorithm as described in the 2010 paper "Enhanced sampling and
applications in protein folding in explicit solvent" by Zhang and Ma.

In this paper <http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2905458/>, the
updated temperatures computed by the adaptive tempering algorithm are said
to be realized by force rescaling. This rescaling changes the heights of
potential energy barriers, allowing for enhanced conformational sampling
without a change in the system temperature (i.e. the kinetic energy as
controlled by the thermostat does not vary outside of its maxwell
distribution at the thermostat value). Consider this quote from the paper's
"Application in Folding Small Proteins" section for support:

 "the scaled potential energy transforms a canonical ensemble at T0 to
another temperature T. On the other hand, the thermostat temperature T0,
which controls the kinetic energy, is unaffected and can be maintained at a
fixed value T0=480 K"

However, the NAMD implementation of the tempering algorithm calls the
system thermostats, either Langevin or velocity rescaling (only two options
compatible with adaptive tempering), to use the updated temperatures. In
such case, the thermostat evolves over time according to the adaptive
temperatures, rendering the kinetic energy variable. The following snippet
from the NAMD controller class reference illustrates this:

01634 if (simParams
<http://www.ks.uiuc.edu/Research/namd/doxygen/classController.html#p84>->langevinOn
<http://www.ks.uiuc.edu/Research/namd/doxygen/classSimParameters.html#o244>)
01635 adaptTempT
<http://www.ks.uiuc.edu/Research/namd/doxygen/classController.html#p114>
= simParams <http://www.ks.uiuc.edu/Research/namd/doxygen/classController.html#p84>->langevinTemp
<http://www.ks.uiuc.edu/Research/namd/doxygen/classSimParameters.html#o245>;
01636 else if (simParams
<http://www.ks.uiuc.edu/Research/namd/doxygen/classController.html#p84>->rescaleFreq
<http://www.ks.uiuc.edu/Research/namd/doxygen/classSimParameters.html#o260>
> 0)
01637 adaptTempT
<http://www.ks.uiuc.edu/Research/namd/doxygen/classController.html#p114>
= simParams <http://www.ks.uiuc.edu/Research/namd/doxygen/classController.html#p84>->rescaleTemp
<http://www.ks.uiuc.edu/Research/namd/doxygen/classSimParameters.html#o261>;

Am I correct to conclude that the NAMD implementation of the tempering
algorithm is different from that described in the paper (granting
conceptual similarity)? If so, is the NAMD implementation's coupling of the
stochastic equation used to guide the adaptive temperature walk with the
stochastic equations that govern the thermostats (Langevin/velocity
rescaling) problematic for sampling? I ask because I've noticed extremely
rapid and large changes in the system temperature (over 100K in under 1000
fs) for simulations in which I have tempering turned on.

I'd appreciate any insight on this matter.

Thanks and regards,

Sachin Natesh

This archive was generated by hypermail 2.1.6 : Tue Dec 27 2016 - 23:21:20 CST