- Contemporary messages sorted: [ by date ] [ by thread ] [ by subject ] [ by author ] [ by messages with attachments ]

From: Jason Swails <jason.swails.gmail.com>

Date: Wed, 30 Jun 2021 23:02:20 -0400

On Tue, Jun 22, 2021 at 9:30 AM Richard Kullmann <

Richard.Kullmann.mpikg.mpg.de> wrote:

*> Hi all,
*

*>
*

*> I am currently trying to speed up my simulations using Hamiltonian
*

*> Replica Exchange MD. Some papers suggested to use multiple copies of one
*

*> replica, for example 5 replicas with the physical parameters, in order
*

*> to enhance the sampling. Has this already been implemented in AMBER
*

*> somehow?
*

*> As far as I understand, having multiple copies with similar parameters
*

*> will just result in 100% exchange rates between these states, therefore
*

*> not really enhancing anything.
*

*>
*

If all you do is a REMD run with a set of identical replicas, then it is

exactly indistinguishable from running the same number of independent

simulations - you get exactly the same ensemble of structures (albeit mixed

through different trajectory files if using H-REMD). However, since doing

REMD has some measurable overhead, your total *effective* efficiency in

this scenario would go down relative to independent simulations.

You are correct that you'll get a 100% acceptance rate, which makes

complete sense: each ensemble is identical, so every configuration

generated by every simulation is an equally valid member of any of the

running ensembles.

However, if you have a situation similar to one where you have, say, 10

total replicas with 5 being run at some "base" Hamiltonian (that you care

the most about), then it's possible you may achieve more efficient sampling

that you care about than if all 10 were run at different Hamiltonians (with

only one being "of interest").

You can certainly *do* this within Amber's H-REMD capabilities by simply

using the same topology and input files for multiple replicas. This is,

roughly speaking, equivalent to doing asynchronous REMD (that is, where

each replica can run a different number of steps between exchange attempts)

where the duplicated replica is run "faster" (in the example above, that

"base" Hamiltonian would be effectively running 5x faster than the other

replicas, which would all be running at the same speed).

A possible real-world example may involve Thermodynamic Integration (where

windows with 0 < lambda < 1 represent fictitious, alchemical states and

lambda = 0/1 represent physical end-states), where you may have 20 total

replicas with 7 identical replicas at both lambda=0 and lambda=1 and the 6

remaining replicas with lambda spaced between 0 and 1. We can compare this

to "traditional" H-REMD TI where each lambda is represented by 1 replica:

In the traditional approach, there would be 8 total replicas, of which 2

would represent "real" states (so 25% of your total sampling would be of

physically relevant, non-alchemical systems). By contrast, in the

duplicate-replica approach I described, each physical end-point would be

simulated with 7 replicas, so 14 of the 20 total replicas would represent

"real" states (so 70% of your total sampling would be of physically

relevant, non-alchemical systems).

I'd be surprised if the approach I described had much of an effect (I'm not

current on the literature right now). But I could certainly come up with

explanations as to why it *could* have an effect, and it certainly may have

a much bigger effect than I intuitively suspect.

HTH,

Jason

Date: Wed, 30 Jun 2021 23:02:20 -0400

On Tue, Jun 22, 2021 at 9:30 AM Richard Kullmann <

Richard.Kullmann.mpikg.mpg.de> wrote:

If all you do is a REMD run with a set of identical replicas, then it is

exactly indistinguishable from running the same number of independent

simulations - you get exactly the same ensemble of structures (albeit mixed

through different trajectory files if using H-REMD). However, since doing

REMD has some measurable overhead, your total *effective* efficiency in

this scenario would go down relative to independent simulations.

You are correct that you'll get a 100% acceptance rate, which makes

complete sense: each ensemble is identical, so every configuration

generated by every simulation is an equally valid member of any of the

running ensembles.

However, if you have a situation similar to one where you have, say, 10

total replicas with 5 being run at some "base" Hamiltonian (that you care

the most about), then it's possible you may achieve more efficient sampling

that you care about than if all 10 were run at different Hamiltonians (with

only one being "of interest").

You can certainly *do* this within Amber's H-REMD capabilities by simply

using the same topology and input files for multiple replicas. This is,

roughly speaking, equivalent to doing asynchronous REMD (that is, where

each replica can run a different number of steps between exchange attempts)

where the duplicated replica is run "faster" (in the example above, that

"base" Hamiltonian would be effectively running 5x faster than the other

replicas, which would all be running at the same speed).

A possible real-world example may involve Thermodynamic Integration (where

windows with 0 < lambda < 1 represent fictitious, alchemical states and

lambda = 0/1 represent physical end-states), where you may have 20 total

replicas with 7 identical replicas at both lambda=0 and lambda=1 and the 6

remaining replicas with lambda spaced between 0 and 1. We can compare this

to "traditional" H-REMD TI where each lambda is represented by 1 replica:

In the traditional approach, there would be 8 total replicas, of which 2

would represent "real" states (so 25% of your total sampling would be of

physically relevant, non-alchemical systems). By contrast, in the

duplicate-replica approach I described, each physical end-point would be

simulated with 7 replicas, so 14 of the 20 total replicas would represent

"real" states (so 70% of your total sampling would be of physically

relevant, non-alchemical systems).

I'd be surprised if the approach I described had much of an effect (I'm not

current on the literature right now). But I could certainly come up with

explanations as to why it *could* have an effect, and it certainly may have

a much bigger effect than I intuitively suspect.

HTH,

Jason

-- Jason M. Swails _______________________________________________ AMBER mailing list AMBER.ambermd.org http://lists.ambermd.org/mailman/listinfo/amberReceived on Wed Jun 30 2021 - 20:30:02 PDT

Custom Search