[gmx-users] dgdl convergence in free energy calculation

[friendli]

Dear all,

I am calculating the mutation free energies using TI method for a small
globulin protein which contains about 40 residues.
I meet convergence problems in the FE calculations:
(1) when simulations start with different velocities for the some lambda
points, the variance of dGdlamba is big and comparable with ddG (several
kJ/mol). That makes the calculation not well reproducible.
(2) the error estimates using block averaging method is also large. The
error is comparable calculated ddG ( ~4 kJ/mol ). Thus the results seems not
very reliable.

By reading, I understand this is due to inefficient sampling for mutations
in proteins.
I attach one of my dgdl plots for LJ mutation(W->A) step (charges for
mutated atoms are removed in previous step). I try to increase the dgdl data
collection time to 2ns, which is relative long compared to time used in
protein mutation in the references. But the large variance in lambda=0.7,
0.8, 0.9 seems not get improved. The error is also not well reduced.
The follwing are the dG results and error in kJ/mol.
dG(500ps)          =  1.01       error = 3.92
dG(500ps_run1) =  -0.40       error = 3.83
dG(1ns)             =   2.21       error = 3.67
dG(2ns)             =   2.44       error = 3.22
dG(2ns_run1)     =  -1.02       error = 2.48



Can anybody give some suggestions on how to improve the convergence and
reduce the statistical error?

I appreciate anyone with free energy calculation experience can give some
advice.



Thanks

Qiang
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20090405/b071a0db/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: dgdl.jpg
Type: image/jpeg
Size: 32729 bytes
Desc: not available
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20090405/b071a0db/attachment.jpg>