alchemical path in Gromacs. The same procedure is carried out for
the ligand in solution. In this case, however, the restraints are
turned on after the decoupling of the ligand intermolecular inter-
actions. However, as discussed in Subheading2.3.1, when using
the set of restraints proposed by Boresch et al. [32] and described
here, these restraining steps do not need to be simulated since this
free energy difference can be calculated analytically.
In general, it is difficult to know a priori how many windows
should be run and for how long for a specific system that has not
been tested before. Test runs are then useful in order to have an
idea about phase space overlap, convergence, and precision of the
calculations; we will discuss how one can assess these in Subheading
4.6. In practice, the computational resources available play a role as
well: the more sampling (i.e., simulation length or number of
repeat runs) the better, but if computational efficiency is required
(either due to large scale calculations, or limited resources) it
becomes even more important to test the calculations in order to
find the setup that maximizes precision for the resources available.
To provide an idea of what the calculations might involve, for
systems containing drug-like ligands binding to small (~110 resi-
dues) and fairly rigid proteins, we typically employed 42 windows
for the complex and 31 for the ligand simulations, each lasting
10–15 ns.
Ideally, theλvalues for the intermediate states should be spaced
in such a way that the statistical uncertainty in the free energy
difference between neighboring states is equal, as this results in
the lowest variance path [53]. In practice, this might be hard to
achieve; yet, if the uncertainty of theΔGbetween two states is
particularly large, it is evident that more windows or tighter spacing
is needed. Whether tighter lambda spacing is necessary can also be
visually evaluated by looking at the plot ofh∂U/∂λiversusλused
for TI: where the slope of the curve changes more rapidly more
windows are needed (an example is shown Fig.4). The windows
used for decoupling the ligand charges can generally be spaced
linearly, e.g.,λcoul ¼[0.0, 0.2, 0.4, 0.6, 0.8, 1.0]. Using a LJ
soft-core potential,λvdwcan also be spaced linearly to start with,
and adjusted as needed. However, more windows are typically
needed for the LJ than for the charge decoupling. For the
restraints, tighter lambda spacing is typically needed when inserting
the harmonic potentials [33]. For instance, withλrestr¼0 being the
state without restraints andλrestr¼1 the state with fully coupled
restraints, a possible spacing isλrestr¼[0.0, 0.01, 0.025, 0.05,
0.075, 0.1, 0.15, 0.2, 0.35, 0.5, 0.75, 1.0].
While the number, spacing, and length of the windows will
need to be adapted to the specific system of interest and the desired
precision, there are a few rules that should always be followed.
Some of these have already been mentioned or alluded to in the214 Matteo Aldeghi et al.