[gmx-users] Large dV/dlambda value at lambda = 0

[Robert Johnson]

Hello everyone,
I'm computing the relative binding free energies of DNA bases on a
nanotube using the protocol listed here:
http://www.dillgroup.ucsf.edu/group/wiki/index.php/Free_Energy:_Tutorial
I'm using the following steps:
1. Turn off all charges on base
2. Transform LJ parameters
3. Turn on all charges

Because the transformation (from G to A for example) involves changing
every single atom charge, I decided to first reduce all charges to
zero. The problem I'm experiencing occurs at small lambda values
during step 2 (transforming LJ parameters) of the above routine. I
obtain large dV/dlambda values of about 2E3 for lambda 0.05 or less.
The rest of the curve contains dV/dlambda of about 6E2. I don't think
this is due to any equilibration issues. Each lambda point is run for
15 ns and I incur a small error of ~1E-1 for each dV/dl value. The
error is a little bit larger (~1) for the lambda=0 point. I believe
the topology and simulation protocol is correct. I have already run
calculations for the absolute binding free energy of this system and I
get good agreement for steps 1 and 3. However, it seems that step 2 is
problematic. If I can't get rid of the large dV/dlambda value, how
should I handle that? Is this large value an artifact? Or should I
assume that the "real" dV/dl curve can be extrapolated to a value that
is more reasonable and in better agreement with the rest of the curve?
If anyone has some suggestions on how to alleviate this problem, I'd
appreciate reading them.

Thanks,
Bob