metadynamics and PMF
Michael Brunsteiner
Dear all,
I was trying to calculate a PMF for a range of colloidal molecules
through metadynamics
where the only collective variable is simply the distance between the
two molecules.
using Gromacs+PLUMED.
plumed.dat:
PRINT W_STRIDE 100
HILLS HEIGHT 0.5 W_STRIDE 100
DISTANCE LIST 241 482 SIGMA 1.0
ENDMETA
I varied the hills height parameter, but irrespective of that
there always seems to be a couple of issues:
1) using sum_hills
"sum_hills -file HILLS -out fes.dat -ndim 1 -ndw 1 -kt 2.5 -stride 10000
-fix 1.2 4.0"
i get identical results no matter how I set "-kt"
Gromacs has energy units of kJ/mol (2.5) while the default value
in sum_hills seems to be kcal/mol (0.6) and the latter seems to be used
not matter which -kt command line argument i give.
I looked at the soure ... can it be that f90 is not case sensitve? because,
if it is, then KT is determined by the command line but not kt, which
seems to be used.
in the calculations...
2) the longer the simulation time the more "Inf" values i see in
fes.dat, the output
of sum_hills... in fact it looks as though the PMF doesn't get
normalized properly
as more and more of the Gaussians in HILLS are summed up, there seem to some
limits that are hard coded in sum_hills leading to these inf values ...
3) the free energy profile in fes.dat is always negative as the the
collective variable (the
distance between the two colloids) gets very small. However, since at
small distances the molecules
begin to overlap, and the energy/repulsion becomes very high and
positive I'd expect the same
for the calculated PMF, but I generally see that at the smallest
distance the free energy is still
lower (more negative) than when the molecules are well separated. Since
i simulate at T=300
I would expect the molecules to sample at least some of the phase space
that corresponds
to a small distance with considerable repulsion, but this is not
reflected in the free energy output.
4) I calulated the PMF for the same system without plumed, using
umbrella sampling
(pull code) and WHAM in Gromacs, where the PMF calculation typically
converges after about
(or less than) 8 nano seconds total simulation time. I wonder if anybody
has compared
the efficiency of metadynamics versus umbrella sampling+WHAM since, as
far as i can see
from my preliminary results with metadynamics (PLUMED+gmx) i don't seem
to get anywhere
near a converged PMF in the same total simulation time. I would expect
metadynamics to
be more efficient, as here the bias force is updated on the fly in real
time, a bit similar
to adaptive umbrella sampling. So I guess my simulation setup is
sub-optimal in some way,
but I am not sure where to start changing that.
thanks in advance for any help/advice!
cheers,
Michael
Giovanni Bussi
This behavior is correct. kt is only used when you integrate over one
or more CVs. In your case, simulation is done with 1 cv and you plot 1
cv, so no reduction is done. Thus, units of fes.dat are kj/mol
> 2) the longer the simulation time the more "Inf" values i see in fes.dat,
> the output
> of sum_hills... in fact it looks as though the PMF doesn't get normalized
> properly
> as more and more of the Gaussians in HILLS are summed up, there seem to some
> limits that are hard coded in sum_hills leading to these inf values ...
This looks strange to me... How many hills do you have? (wc -l HILLS).
Also, it seems that your hills are quite big for a very long 1d
simulation.
>
> 3) the free energy profile in fes.dat is always negative as the the
> collective variable (the
> distance between the two colloids) gets very small. However, since at small
> distances the molecules
> begin to overlap, and the energy/repulsion becomes very high and positive
> I'd expect the same
> for the calculated PMF, but I generally see that at the smallest distance
> the free energy is still
> lower (more negative) than when the molecules are well separated. Since i
> simulate at T=300
> I would expect the molecules to sample at least some of the phase space that
> corresponds
> to a small distance with considerable repulsion, but this is not reflected
> in the free energy output.
Absolute free energy has no meaning in this context. Only differences
matter. F(s) is always (conventionally) negative in metadynamics,
since it is obtained from the negative of the bias potential.
>
> 4) I calulated the PMF for the same system without plumed, using umbrella
> sampling
> (pull code) and WHAM in Gromacs, where the PMF calculation typically
> converges after about
> (or less than) 8 nano seconds total simulation time. I wonder if anybody has
> compared
> the efficiency of metadynamics versus umbrella sampling+WHAM since, as far
> as i can see
> from my preliminary results with metadynamics (PLUMED+gmx) i don't seem to
> get anywhere
> near a converged PMF in the same total simulation time. I would expect
> metadynamics to
> be more efficient, as here the bias force is updated on the fly in real
> time, a bit similar
> to adaptive umbrella sampling. So I guess my simulation setup is sub-optimal
> in some way,
> but I am not sure where to start changing that.
In general, if you use just one cv there is not a big difference
between these methods, even if people have their "favorite" method and
usually stick to it. Metadynamics is really superior in multiple
dimensionality, since it allows to have an estimate free energy just
visiting the relevant part of the landscape.
Anyway, results should agree, otherwise there is something wrong in
your setup/input files/analysis.
Giovanni
>
> thanks in advance for any help/advice!
>
> cheers,
> Michael
>
>
> --
> You received this message because you are subscribed to the Google Groups
> "PLUMED users" group.
> To post to this group, send email to plumed...@googlegroups.com.
> To unsubscribe from this group, send email to
> plumed-users...@googlegroups.com.
> For more options, visit this group at
> http://groups.google.com/group/plumed-users?hl=en.
>
>
Giovanni Bussi
It also seem that SIGMA is too large. Is is really that you want a 1
nm resolution?
Giovanni
Massimiliano Bonomi
> of sum_hills... in fact it looks as though the PMF doesn't get normalized properly
> as more and more of the Gaussians in HILLS are summed up, there seem to some
> limits that are hard coded in sum_hills leading to these inf values ...
Hi Michael!
You may want to give a look at this post:
http://groups.google.com/group/plumed-users/browse_thread/thread/dab90acced885466?hl=en
Best,
Max