A Separating Framework for Increasing the Timestep in Molecular Dynamics
Recent work on our LIN framework for molecular dynamics is described. This
separating framework solves the Langevin equations of motion in two steps:
linearization and correction. The linearized equations of
motion are solved numerically by an iterative technique whose cost is dominated
by sparse-Hessian/vector products; the resulting 'harmonic' solution is
then corrected by an implicit integration step, which requires minimization
of a nonlinear function. Computational efficiency is achieved by sparse
matrix techniques, adaptive timestep selection, and fast minimization.
Applications to the model systems of alanine dipeptide and BPTI demonstrate
very similar trajectories with LIN at moderately large timesteps
(15 fs for BPTI and up to 30 fs for the dipeptide) in comparison to traditional
MD with much smaller timesteps (0.5 fs). This agreement validates the LIN
approach at these timesteps. The net computational gain is modest for BPTI
(30% reduction in total time) but is expected to increase with system size.
Moreover, examination of the range of validity of the harmonic approximation
led also to development of a related method termed LN, which includes LIN's
linearization, but not correction, step. LN at a timestep of 5 fs also
shows very good agreement with traditional molecular dynamics and already
for BPTI gives a speedup factor of 4 (2 for the dipeptide). Speedup will
only increase with size, and the new LN variant can be readily implemented
in general biomolecular dynamics programs. This unexpected windfall in
computational performance illustrates the value of developing novel approaches
(e.g., based on normal modes) that might initially appear not practical
for macromolecules.
Click to go back to the publication list
Webpage design by Igor Zilberman