Removing fastest degrees of freedom

WARNING -- This information has been culled from some gmx-users posts in October and November 2008. This information may be misleading or entirely inaccurate.

In gromacs-4.0, pdb2gmx -vsite aromatic leads to a fatal error during mdrun:

Fatal error: Virtual site atom 29 is part of a charge group of only virtual sites, but its first constructing atom (26) is part
of a different charge group, this combination is not allowed.

This error message is also valid for older versions. It can not be guaranteed that the virtual sites in one charge group will have the same periodic shift. But before 4.0 there was no check for this.

Using -vsite aromatics are therefore not recommended and it is possible that this option will be removed from pdb2gmx, since you do not gain so much compared with dt=4 or 5 fs with -vsite h.

If you really want to use the -vsite aromatics, you can put all atoms in the aromatic groups in separate charge groups, but that will make the simulation slightly slower. Further, putting all of the atoms in each aromatic side chain into their own charge group could slightly change the equilibrium properties of a peptide or protein since aromatic residues would then be more strongly attracting near the cutoff distance than they are when assigned the regular charge groups of a force field. Most force fields (except GROMOS) have probably been parametrized without charge groups. Therefore having smaller charge groups is always more accurate, unless you have a buffer region, which is now possible in 4.0. (This all assumes that you use PME and not cut-off or reaction-field for electrostatics). However, this is probably not a serious issue.

One might try: pdb2gmx -vsite h, with:

dt = 0.004
constraints =  all-bonds
constraint_algorithm =  lincs
lincs-iter =  1
lincs-order =  6

Don't forget to change nstlist based on the largest end-to-end distance of any charge group and the largest possible diffusion between neighbour list updates.

Using 5 fs without heavy hydrogens will not cause stability problems. It might cause slightly less accurate integration. But these integration errors are usually much smaller than other errors, such as cut-off errors. It is all a matter of what accuracy you want. If you want to be sure, you can run simulations with a 4 and 5 fs timestep (do not forget to modify nstlist as well, so the neighborlist update frequency stays constant) and then check if the properties you are interested in change. For many properties this might not be feasible, since you can not converge the properties within reasonable time.

In some users experience 5 fs works fine and they would not hesitate to use it for large systems, where one can anyhow not sample all conformational space. One user is currently mainly doing accurate calculations of thermodynamic quantities, for these they use 4 fs.