Multiple time step algorithms for molecular dynamics simulations of proteins: How good are they?

We evaluate several multiple time step (MTS) molecular dynamics (MD) methods with respect to their suitability for protein dynamics simulations. In contrast to the usual check of conservation of total energy or comparisons of trajectory details, we chose a problem‐oriented approach and selected a set of relevant observables computed from extended test simulations. We define relevance of observables with respect to their role in the description of protein function. Accordingly, the use of quantities that exhibit chaotic behavior, like trajectory details, is shown to be inappropriate for the sake of the evaluation of methods. The accuracy of a cutoff method and of six MTS methods is evaluated, which differ in their treatment of the computationally crucial long‐ranged Coulomb interactions. For each of the observables considered, the size of purely statistical fluctuations is determined to allow identification of algorithmic artifacts. The obtained ranking of the considered MD methods differs significantly from that obtained by the usual measures of algorithmic accuracy. One particular distance class method, DC‐1d, is shown to be clearly superior in that no algorithmic artifacts were detected. © 1998 John Wiley & Sons, Inc. J Comput Chem 19: 1534–1552, 1998