Talk:Best Practices
We may want to expand the following and include it on the main page
Additional potential guidelines for best practices:
- Beware changes in net charge; don't assume that these will automatically be handled correctly. Add references to support this; hopefully figure out the solution/solutions and explain. Relevant work is by Hunenberger, recently, in JCP, and also older work by Hummer.
- Some discussion of nonequilibrium methods (slow/fast growth): So far the evidence is that (a) slow growth is in general problematic, because of things like Hamiltonian lag, and can introduce hysteresis, etc. Really, as Jarzynski has more recently shown, one would need to average over an ensemble of nonequilibrium simulations in order to get correct free energy differences. (b) fast growth needs lots of realizations in the fast limit. And several recent studies (i.e. the calculating zeros paper by Oostenbrink and van Gunsteren, and several others, I think from Geissler and from Ytreberg(?)) indicate it is in many cases less efficient than equilibrium methods. So our recommendation is to use equilibrium methods.
Other material that needs to be fleshed out and have references incorporated for the main page
Simulation methodology
Free energy calculations typically involve calculating free energy differences that are relatively small compared to the total potential energy of a system. As a consequence, certain simulation parameters which may be unimportant for “typical” molecular dynamics simulations (because they change the total potential energy by such a small percentage, for example) can be tremendously important in free energy calculations. Thus, it is important to think carefully about simulation settings used for free energy calculations. Here are some examples of issues we have found to be important:
- PME parameters: When using lattice-sums (Ewald, PME, PPME, and variants) to handle long range electrostatics, setting details for these can substantially affect computed free energies. In particular, one needs to ensure that the settings chosen give electrostatic interaction energies accurate to very small fractions of a kcal/mol. If this condition is not met, computed free energies can be wrong – sometimes even by several kcal/mol. To test this, one can compute accurate electrostatics energies for a set of simulation snapshots by evaluating their energies using very long electrostatics cutoffs, then re-evaluate energies using shorter cutoffs and lattice-sum electrostatics. Settings for lattice-sum electrostatics and electrostatic cutoffs should be chosen so that the total potential energy from lattice sum is accurate (compared to that evaluated with a very long cutoff) to a very small fraction of a kcal/mol (less than the desired uncertainty in the computed free energies). Some work has found that default settings for some simulation packages may lead to errors in free energies of up to several kcal/mol [Mobley].
- Thermostat choice: The choice of thermostat (for constant temperature simulations) can be quite important, especially in absolute free energy calculations. Many thermostats perform well when a system has a sufficiently large number of degrees of freedom, but this is not an adequate criteria for absolute free energy calculations, as at the end state in these calculations, a portion of the system typically does not interact with its environment. This portion – i.e. a small molecule – may have comparatively few degrees of freedom. Hence, simulations done with thermostats that do not sample from the correct distribution of kinetic energies even in the limit of few degrees of freedom may exhibit problems such as non-ergodicity near these end states (Shirts, Mobley unpub, Villa and Mark 2001). Thus it is important to use thermostats (like Andersen’s thermostat (ref) or a chain of Nose-Hoover thermostats (ref)) or dynamics (like Langevin dynamics) which are known to sample from the correct ensemble under such circumstances.
Of course, this is by no means an exhaustive list, and these issues may be simulation-package dependent. We encourage further exploration in this area, and suggest that future work with free energy calculations should at minimum perform the above checks. Preferably, users should get used to testing different settings in their simulation packages to ensure that these are set to obtain sufficient accuracy for free energy calculations.
In addition to these issues which can be particularly important for free energy calculations, there are a lot of choices to be made in setting up your system for simulation that also need to be made in a sensible way. See our page on MolecularDynamicsSetup for some of the issues to consider, such as protonation states, parameters for proteins and small molecules, missing residues in protein structures, simulation box sizes, and so on.