The Generalized ParRep algorithm: developing an OpenMM based implementation for studying biochemical systems

Florent Hedin
École des Ponts ParisTech

The Parallel Replica (ParRep) algorithm, first introduced by Arthur Voter [1] and later mathematically formalized [2], relies on the convergence of stochastic dynamics (e.g. Langevin Dynamics) to a Quasi Stationary Distribution (QSD) when the dynamic evolves within a metastable state.

Some properties, such as the exit time, may therefore be estimated accurately as they follow an exponentially decaying law.
It was shown [1,2,3] that, in the case where N replica of the system are running in parallel, one can estimate the average exit time by multiplying the first observed exit time over N replica by N, if the replica were initially properly decorrelated and converged to the QSD. Therefore a speedup factor of N can theoretically be achieved.

My task is to develop a robust, flexible and fast implementation that one can apply to any chemical or biological systems.
For that the current implementation relies on the openMM toolkit [4,5] for performing the molecular dynamics.
Fast and efficient parallelization is achieved by using C++ and OpenMPI.
First successful applications consisted in the study of the mean isomerization time of alanine dipeptide, and mean dissociation time for the protein-ligand system FKPB-DMSO.

[1] A. F. Voter, Phys. Rev. B 57 (1998)
[2] C. Le Bris, T. Lelièvre, M. Luskin and D. Perez, Monte Carlo Methods Appl. 18 (2012)
[3] A. Binder, T. Lelièvre, G. Simpson, J. Comput. Phys. 284 (2015)
[5] P. Eastman et al, DOI: 10.1101/091801 (2016)

Back to Complex High-Dimensional Energy Landscapes