[gmx-developers] re: Energy and Foreign Lambdas and Hamiltonian Replica Exchange

[Matt Wyczalkowski]

On Feb 15, 2008 4:56 AM, Erik Lindahl <lindahl at cbr.su.se> wrote:
> Hi,
>
> My own server is unfortunately down since it was hacked last week and we
> were anyway about to move from Joomla to Dekiwiki, but I've uploaded a copy
> to the ftp site. Have a look in
>
> ftp://ftp.gromacs.org/pub/tmp/multilambda/
>
> Not sure if it's useful for your stuff, and since I haven't gotten much
> feedback for it yet you shouldn't use it for production purposes without
> significant testing. You've been warned, have fun ;-)

I've completed a first round of testing.  To get the current version
(VERSION 3.3.99_development_20071120) to compile, I added the
following 3 lines to the file include/types/inputrec.h,
  #define DECOUPLING_VDW                 (1<<0)
  #define DECOUPLING_COULOMB             (1<<1)
  #define DECOUPLING_BONDED              (1<<2)
This was a guess, but seems reasonable given what the constants do and
I'm not using that functionality anyhow.  After this (and commenting
out a for-humans-only note in the same file), compilation proceeded
without a hitch.

Reading through a generated MDP file gives good insight into the
functionality of the code, and as an initial test I tried to make sure
that foreign energies (which are specified by the deltaE-lambda
parameter) match what I'd expect.

The system I ran consisted of a single neutral amino acid residue in a
1.1nm box with 45 or so TIP3P waters; all cutoffs are 0.5nm.  The
parameter lambda controls only the charge of all the atoms in the
residue, so that at l=0 it is uncharged, and at l=1 the charges are
the OPLS-AA charges.  For the first round of tests, I simply evaluated
the energies for the initial configuration of the system. What I found
was that the native energy (as reported in the log file) differs
depending on whether deltaE-lambda is defined (as deltaE-lambda=1, in
this case) or not, and that this behavior is dependent on the
electrostatic implementation used.  Of course, the native energy of a
system should not change due to a foreign energy evaluation.

Ewald and PPPM are explicitly not supported, and fail politely when
free energy calculations are requested.  The Reaction Field energy
term (RF excl.) differed by an order of magnitude depending on whether
deltaE-lambda was defined.  PME showed similar errors (but of a
smaller magnitude).  Finally, simulations with Cutoff electrostatics
were consistent whether or not deltaE-lambda was defined.

Next I looked at the foreign energies (at various lambda), and checked
them against the native energy of the system.   The foreign energies
in the cutoff electrostatics system seem consistent with what I'd
expect: in a lambda=1 simulation foreign energy at lambda=1 matches
the native energy.  Also, for a lambda=0.5 simulation, the lambda=1
foreign energy is consistent with the native energy of the lambda=1
simulation, as one would expect, and vice versa.

In summary, both the Reaction Field and PME electrostatics seem to
have broken foreign energy calculations, whereas the Cutoff
implementation seems OK according to these very preliminary tests.

Regards,

Matt Wyczalkowski

-- 
Matt Wyczalkowski
Doctoral Candidate, Biomedical Engineering
Pappu Lab: http://lima.wustl.edu
Washington University in St. Louis