Re: convergence criterion of free energy perturbation (FEP) simulations for pKa calculation

From: Brian Radak (
Date: Thu Oct 05 2017 - 15:30:28 CDT

Does your transformation also include bonded terms? You will not compute
a relative ionization free energy if you do not include those terms,
although you can, in principle, compute a difference in relative
solvation free energies by doing the corresponding calculation in the
gas phase. I don't know your reasons for running your simulation, but I
feel that I should point this out -- there are new keywords as of NAMD
2.11 for doing this (see the user guide).

Agreement between the forward and backward FEP calculations is only a
necessary condition for FEP convergence -- it is neither a sufficient
condition for convergence nor a meaningful condition for other
estimators, such as BAR.

Another necessary, but still not sufficient, condition for convergence
might be found in the long time behavior of the BAR estimate of the
cumulative data (i.e. use the first 10 ns, then the first 20 ns, etc.
and see if the answer changes as the simulation gets longer).

The forward and backward construction loses all meaning when analyzing
with thermodynamic integration, for the same reason as it does for BAR.


On 10/05/2017 02:36 PM, Sadegh Faramarzi Ganjabad wrote:
> Hello,
> I am running a test FEP simulation to calculate free energy of the
> ionization of an amino acid inside a membrane protein. Here is the FEP
> part of my namd config.
> source                  ../tools/fep.tcl
> alch                    on
> alchType                FEP
> alchFile                all-ion-beta.pdb
> alchCol                 B
> alchOutFile             forward-shift-pr.fepout
> alchOutFreq             10
> alchBondDecouple        on
> alchBondLambdaEnd       1.0
> alchVdwLambdaEnd        1.0
> alchElecLambdaStart     0.5   #soft potential
> alchVdWShiftCoeff       6.0
> #alchDecouple            off
> alchEquilSteps          100
> set numSteps            500
> runFEPmin 0.0 1.0 0.0625 10000 1000 310
> Then I used ParseFEP of VMD. The difference between DeltaG of the
> forward and backward transformation is about 5 kcal/mol for lambda=0
> to 0.5, and it become almost 0 for lambda = 0.6 to 1. I should
> probably increase simulation time for each lambda from 10,000 to say
> 10,000,000 until the plots of forward and backward lie on top of each
> other. My question is that is this the only criterion for convergence
> of FEP calculation? How can I make sure my simulation is long enough?
> Also, if I use thermodynamic integration method for free energy
> calculations, what would be convergence criteria for that?
> Thanks,
> Sadegh

Brian Radak
Postdoctoral Appointee
Leadership Computing Facility
Argonne National Laboratory
9700 South Cass Avenue, Bldg. 240
Argonne, IL 60439-4854
(630) 252-8643

This archive was generated by hypermail 2.1.6 : Mon Dec 31 2018 - 23:20:38 CST