[gmx-developers] More about the Ekin issue.

[Berk Hess]

Berk Hess wrote:

>
>>> But do you also run the REMD in NpT?
>>> Then you have a pV term which you probably did not account for in 
>>> the exchange.
>>> If you do account for the pV term it is around 10x as larger as the 
>>> energy difference
>>> and will therefore ruin the efficiency (and maybe accuracy) of your 
>>> REMD.
>>>   
>>
>> According to Xavier's mdp file he does use NPT simulation, the question
>> is which REMD he uses...
>> I'm not sure whether including the pV term is bad, it make the
>> acceptance ratio go to 40%, but the overall number of exchanges also
>> depends on the frequency with which you try to exchange. The guys that
>> make all these algorithms don't usually test it on proteins, but on
>> simple model systems.
>>  
>>
> The acceptance ratio does not say anything about efficiency.
> The pV term is roughly 10x larger the energy term, so the energy 
> difference
> has a negligible influence on the exchange probability.
> The exchange probability is almost fully determined by pV.
> Fluctuations in the solvent probably contribute most to pV.
> So you just exchange based on fluctuations in the solvent
> and not on the conformation of the solute.
> I have no idea how efficient REMD based (mainly) on the pV criterion is.
>
> Another issue is that the pV distribution with Berendsen pressure
> coupling is quite off for small systems. For 400 waters:
> Berendsen: pV RMSD 375 kJ/mol
> Parrinello-Rahman: pV RMSD 490 kJ/mol
>
Now I thought a bit more I think the huge fluctuations are the problem.
The average pV for 400 waters should be around 12 kJ/mol.
If there were no large fluctuations the exchange criterion would produce
proper and efficient REMD sampling. An increased temperature might
increase the average pV by 0.1 kJ/mol.
But the enourmous fluctuations will dominate the exchanges and cause
all exchanges to be random. If you attempt to exchange infrequently
the normal MD sampling will correct a bit for this, but the ensembles
will probably still be quite incorrect.

(Note that a molecular virial will not help as there are no intramolecular
interactions in water)

Berk.