Network analysis of replica exchange simulations

Abstract

Replica-exchange molecular dynamics (REMD) simulations have become an increasingly popular method for probing the equilibrium properties of biomolecules. REMD executes simulations of same system evaluated under different conditions and periodically attempts to exchange structures at neighboring conditions according to Metropolis Monte Carlo. Our study applies network analysis to assess the impact of the number of replicas on REMD performance and address if graph theory can be a useful tool to aid in the analysis of REMD. Using alanine dipeptide as a model system and a REMD temperature range from 310 K to 500 K, we evaluated the behavior of REMD when 12, 16, 24, 28, 32, and 36 replicas were used. To perform network analysis, we defined a node as a simulation condition and set the distance between nodes based on the empirical exchange failure rate. We computed network centrality metrics and found that average closeness decreases with the number of replicas. Although this result suggests that structural exploration via replica exchange is expected to differ depending on the network, surprisingly the number of states, computed as the number of distinct enthalpy values observed at 310 K, remains the same for all networks. Because this outcome was obtained for a simple simulation system, we will repeat this analysis for more complicated biomolecular systems to validate our conclusions. Overall, our study demonstrates that network analysis is a promising approach to evaluate the technical performance of REMD and should be explored further.