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

[David van der Spoel]

Berk Hess wrote:
> Mark Abraham wrote:
> 
>> David van der Spoel wrote:
>>
>>> Matt Wyczalkowski wrote:
>>>
>>>> Hi,
>>>>
>>>> I'm investigating the feasibility of implementing Hamiltonian 
>>>> Replica Exchange in Gromacs, so that swaps exchange lambda values 
>>>> (rather than temperatures).  To do this, it is necessary to know 
>>>> energies for a given configuration at other lambda values.  This is 
>>>> an issue which was discussed a number of months ago in the context 
>>>> of the Bennett Acceptance Ratio.
>>>>
>>>> Before delving in too far into the code, I wanted to get a feel for 
>>>> the current state of development with respect to energy evaluations 
>>>> at different lambdas, as well as whether Hamiltonian Replica 
>>>> Exchange has been considered and/or implemented.  How significant an 
>>>> undertaking would a generalization of Replica Exchange to allow for 
>>>> lambda swaps be?
>>>>
>>>> Thanks for your time --
>>>>
>>>> Matt
>>>>
>>> I'd say a couple of weeks to get it right. You can learn quite a bit 
>>> from the current REMD implementation, but you'd need extra energy 
>>> evaluations. Check the wiki, to see if this is there already, 
>>> something similar is, anyway.
>>
>>
>> Wasn't there a rework of REMD planned to permit multiple processors 
>> per replica? Has that made it to CVS yet?
>>
>> With the REMD code as it stands in 3.3.1, I'd agree with David's 
>> estimate. That's the kind of time it took me to implement a version of 
>> the hybrid solvation RE method of Okur et al., as described in JCTC 
>> 2(2):420, and that required linking in a PB solver.
>>
>> Mark
> 
> I would say this is very easy, less than a day. But maybe I am overly 
> optimistic,
> since I know the code. This involves no real changes to the code, since 
> you just
> do the acceptance based on the same potential energy of normal REMD.
> There are only some checks that need to be changed.
> 
> But I don't see for what kind of systems such a RE would help.
> Or is this not for free energy differences and do you want to use lambda
> to lower barriers?
> 
IIRC this requires evaluating the energy at a different lambda than the 
one assigned to the processor(s). Maybe someone can post the exchange 
probability criterion.

It is all about sampling of course, and this kind of approach is also 
used in constant pH MD (with implicit solvent) and in Van Gunsterens 
switching between coarse grained and atomistic MD. A more general setup 
for all these REMD variants would be useful therefore. One could also 
envision REMD in two dimensions (e.g. T and lambda).

-- 
David van der Spoel, Ph.D.
Molec. Biophys. group, Dept. of Cell & Molec. Biol., Uppsala University.
Box 596, 75124 Uppsala, Sweden. Phone:	+46184714205. Fax: +4618511755.
spoel at xray.bmc.uu.se	spoel at gromacs.org   http://folding.bmc.uu.se