From: Kenno Vanommeslaeghe (kvanomme_at_rx.umaryland.edu)
Date: Fri Sep 20 2013 - 12:22:39 CDT
As mentioned earlier, there are different kinds of convergence. You seem
to be trying to verify convergence in the sense of your replica exchange
simulation being equilibrated. However, it sounds like the kind of
convergence you are really looking for is convergence of the
conformational ensemble. As also mentioned earlier, fundamentally spoken,
this kind of convergence is impossible to check because of the wide range
in magnitude of conformational barriers. However, pragmatically spoken, it
is often possible to make an educated guess.
Now, to tackle your actual problem at hand. Assuming the phase space
overlap between your replicas is already adequate, I don't think you can
improve conformational sampling significantly by simply increasing the
number of replicas. In your situation, the best way to improve sampling
seems to be to increase the temperature of the highest replica. Of course,
you need to maintain phase space overlap while doing so, which means that,
as the temperature of the highest replica increases, more replicas will be
needed anyway. So I would recommend graphing the histograms of the energy
of the different temperature windows together on one plot, and use that
information to choose a set of 32 temperature windows with the highest
temperature as high as possible, yet low enough so that you can be
confident the histograms will have enough overlap.
Oh yeah, T-remd is indeed known to be pretty painful in explicit solvent,
so your decision to go implicit makes sense. But no convergence check in
the world will give you information on how bad the implicit solvent
approximation is in your particular system. Only comparison with explicit
solvent or experiment will do that.
On 09/20/2013 03:33 AM, Francesco Pietra wrote:
> Hi Niklaus:
> What I am trying to do is defining the conformation of starting and ending
> parts of a homotrimer. These portions, 34aa and 12aa for each subunit in
> our hands do not diffract adequately under x-ray at low temp. So, I guess
> that the energy barriers are low. This is why my hope to get that type of
> T-remd convergence I alluded to. Once modeled, those parts should be
> replaced in the original (by superimposing a rmsd-colvars restrained small
> helical portion).
> Computer time does matter for me indeed, as I am trying to accomplish the
> investigation under a fixed budget. I first tried the 34aa in TIP3 water
> but the request of hardware, in terms of physical nodes, was too high to
> fix "Stray PME grid charges detected" that was raised. Therefore, I
> changed to implicit GB, following other-people idea that most of the
> resources for explicit waster T-remd are wasted with swapping water
> interactions. GB runs fast, however I don't know if is as good as with
> latest adjustment of parameters for amber (C. Simmerling 2013).
> Clearly, I need a test of convergence that is not that provided by
> "show_replicas.vmd". I wonder whether the collection of replicas obtained
> with show_replicas.vmd could in place be elaborated, such as for measuring
> RMSD for the various replicas against the initial unfolded situation. Just
> to have a measure of how much the various replicas differ from one
> another. As I am using 32 replicas (0.7 exchange ratio), the matter is
> rather tricky. With 16 replicas the exchange ratio is still very good
> (0.4) but the computer time saved is minimal (and I pay the same as for 32
> replicas). I hope that convergence is faster with 32 replicas. As a
> biochemist, I never engaged myself in extensive coding, while it would be
> probably not too difficult to adapt show_replicas.vmd to the task. I have
> also thought to move to namd_2.10, where remd is built in, so that it
> should be faster.
> On Thu, Sep 19, 2013 at 7:30 PM, Niklaus Johner <niki.johner_at_gmail.com
> <mailto:niki.johner_at_gmail.com>> wrote:
> Yes each replica samples from the same "extended" phase-space so that
> if you wait long enough, each replica will have visited each
> conformational state often enough to have a converged statistic of
> that state's occupation at the different temperatures and therefore
> you'll be able to say that, as Jason points out, your replicas have
> So fundamentally I agree with you that obtaining this kind of
> convergence would be ideal (if not proof of convergence). I just think
> it's not realistic with our current simulation capabilities, except
> for very small systems with a reasonably small number of degrees of
> freedom, like the alanine-dipeptide (that's why all the papers
> introducing new sampling methods only look at the dipeptide and the
> TRP-cage :-)). So if you get convergence of the probabilities of
> occupancy of the few most sampled states, that would already be an
> achievement and is, in my opinion, enough to give you some confidence
> in the conformations you predict.
> How big is your peptide?
> Obviously looking at different measures of convergence will increase
> your confidence in your sampling. You could look at frequency of
> certain structural elements, convergence of contact maps etc. Split
> the simulation in pieces and compare, compare different replicas...
> A good way to speed up convergence is to start each replica from a
> different configuration. Again, I know it shouldn't matter if you wait
> long enough, but again, I think it's usually not an option to wait
> that long.
> 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 11:53 AM, Francesco Pietra wrote:
>> 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
>> 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.
>> francesco pietra
>> On Thu, Sep 19, 2013 at 4:48 PM, Niklaus Johner
>> <niki.johner_at_gmail.com <mailto: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,
>> 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:
>>> 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