From: Francesco Pietra (chiendarret_at_gmail.com)
Date: Thu Sep 19 2013 - 10:53:58 CDT
Absolutely no. Each individual replica samples from the same pool.
Therefore, if you do not get the same from each individual replica, this
means no convergence. It might be difficult to get convergence but, if not
obtained, it would be non scientific to go to "sortreplicas" to compute
properties at the temperature of your interest.
This is the way I understand T-remd. But I am prone to reeducate myself in
the light of compelling reasoning. This was not the case so far.
cheers
francesco pietra
On Thu, Sep 19, 2013 at 4:48 PM, Niklaus Johner <niki.johner_at_gmail.com>wrote:
> What do you mean by "giving the same answer"? Are you hoping that all your
> replicas will converge to a unique structure? The whole point of T-remd is
> that the replicas switch from one temperature to another, to avoid getting
> stuck in a particular minimum. I think the expectation is that even if you
> could run long enough to have "convergence", meaning that you have the
> correct populations for all the important states, which is more than you
> can hope for, this would be a global convergence of the T-remd and not of
> the individual replicas. Different replicas might explore the phase-space
> around different local minima, but should populate the lower temperatures
> in the simulation according to the relative energies of these minima. So I
> think what you want to test is if the population of states in the lowest
> temperatures is stable. So I would do clustering on different parts of the
> trajectory and compare the clusters of highest occupancy, or something like
> this.
>
> Be aware that you can never really know if a simulation is converged. How
> could you know if you've seen all the important conformations? Maybe the
> most important conformation, with lowest free-energy can be reached from
> your starting configuration only by crossing a very high energy barrier, so
> that you will never see it, and all the measures you can come up with could
> tell you that it's converged. It's probably one of the biggest issues with
> MD.
>
> Good luck,
>
> N.
>
> Niklaus Johner
> Weill Cornell Medical College
> Harel Weinstein Lab
> Department of Physiology and Biophysics
> 1300 York Avenue, Room D-501
> New York, NY 10065
>
>
>
> On Sep 19, 2013, at 2:48 AM, Francesco Pietra wrote:
>
> Hello:
> I posed the same question to the vmd forum but probably is to the namd
> forum pertinent for the question.
>
> Thus, I carried out a short T-remd on alanin with 8 replicas, as provided
> by namd_2.9. I used the final namd-provided folded pdb for comparison. The
> script show_replicas.vmd with these replicas warned "not converged".
>
> With my peptide, 32 replicas, after 370,000 steps still far from
> convergence, I used, for the vmd comparison, the starting unfolded.pdb also
> as a fake folded.pdb. In this case, show_replicas.vmd did not raise any
> warning. Looking also at the code, it seems to me that show_replicas.vmd
> assumes, as a criterion of convergence, the comparison of the various
> replicas with the pdb file that you give as folded peptide in the
> fold.peptide.conf file. That works if your T-remd is just devised to check
> if you are able to reproduce an experimentally defined situation.
>
> Suppose instead that your T-remd is devised to search for the best
> conformation, or cluster of conformations, for an experimentally undefined
> peptide. I can imagine many situations where the experimental approach is
> problematic. Then you need a real criterion of convergence of yor T-remd,
> before starting to examine the (sorted) same-T replicas. Obviously, a
> reliable criterion of convergence is that what you get must be the same
> from each replica, for example that the average structure is the same from
> all replicas.
>
> Therefore, I got the impression that the warning "not converged" raised
> by show_replicas.vmd is misleading. Is any script available to check when
> any replica gives the same answer.
>
> Thanks for advice, even if showing that I am wrong.
>
> francesco pietra
>
>
>
This archive was generated by hypermail 2.1.6 : Wed Dec 31 2014 - 23:21:42 CST