Changes in how Amber handles softcore interactions (August, 2019)
What was changed
In August, 2019, we posted to two significant updates to how Amber
handles softcore interactions in free energy calculations: these were
AmberTools19 update.8 (for sander) and
Amber18 update.15 (for pmemd).
To install these changes, do the following, then recompile:
$AMBERHOME/update_amber --update
Before the update, if one had 1-4 electrostatic or Lennard-Jones
interactions that connected one atom in a softcore region with another
atom in a "fixed" (non-softcore) region, that interaction was made to be
independent of lambda, and hence did not contribute to
dV/dlambda
in the free energy calculation. After the update, such terms are
handled in the "standard" way, that is, the interaction strength is
dependent on lambda
Basically, the error arose from incorrectly using "and" instead of
"or" in a logical statement examining 1-4 interactions. See the changes
to amber18/AmberTools/src/sander/extra_pts.F90 in
AmberTools19 update.8 or to
amber18/src/pmemd/src/extra_pnts_nb14.i in
Amber18 update.15. Thanks to Taisung
Lee, Darrin York, Charlie Lin and David Case for working on the patch.
Our attention to this issue was driven by internal testing (to be
reported elsewhere) and by reports from users, including results of
Loeffler et al., J. Chem. Theory Comput. 14, 5567 (2018).
What the likely effects are
If your free energy calculation doesn't use the softcore option
(ifsc=1), then nothing will change.
If your softcore region doesn't have any covalent connections to a
non-softcore region, then nothing should change.
If your softcore region does have a covalent connection to a
non-softcore region, you should expect to see potentially significant
changes in delta-G values along a particular leg of a thermodynamic
cycle; however, differences between "unified" calculations (where both
LJ and electrostatic energies are removed in a softcore step) and
"split" protocols (where electrostatic energies are treated in separate,
non-softcore steps) should go away.
You may also see changes in delta-delta-G values, i.e. the difference between
two legs of a thermodynamic cycle, and the only value that can be
compared to experiment. There
are circumstances in which the changes may be small:
For fairly rigid systems, where the conformational ensemble
on one leg of the thermodynamic cycle is nearly the same as on the other
leg, the changes imposed by these updates will nearly cancel on the two
legs of the cycle, leading to only small changes in delta-delta-G.
If the softcore region has zero charges, as would be the case
when charges are removed in an earlier step, the remaining 1-4
Lennard-Jones interactions are often quite small, again leading to only
small changes in delta-delta-G.
These updates do not address all possible problems with softcore
calculations. Basically, the Amber codes allow one to choose any
softcore region one likes, without taking into account the nature of the
covalent interactions (if any) between softcore and non-softcore
regions. This actually works pretty well in a large number of cases,
but does not guarantee provide proper end-states for free energy
calculations. Work is ongoing on how to achieve both a reliable and intuitive way
to compute alchemical free energies when some atoms are present in one
end state but not in the other.
In particular, the softcore scheme as currently implemented in Amber
may fail badly if you have a softcore region that is connected to a
common (non-softcore) region by more than one covalent bond. Even with
just one bond to the softcore region, you may see significant effects if
the conformations of the softcore region sampled on one leg of the
thermodynamic cycle are significantly different from those on the other
leg. Problems can also arise if the size of a softcore region is larger
than the cutoff value for non-bonded interactions.
|