An improved TIGER2 implementation for NAMD suitable for the Blue Gene architecture☆
Introduction
Molecular Dynamics simulations are often faced with the challenge of ensuring sufficient sampling of configuration space for the system under consideration. One approach commonly employed to improve this is replica-exchange molecular dynamics (REMD), in which multiple identical copies of the system are simulated concurrently using different temperatures [1] or Hamiltonians [2], [3], [4], [5]. Exchanges are then attempted between adjacent replicas at fixed time intervals and accepted based on the Metropolis criterion [6].
One class of REMD method involves running each replica at a different temperature with attempts to exchange with neighbouring replicas being made at regular intervals [1], [7], referred to herein as Temperature REMD (T-REMD). This method enhances the conformational sampling of the replica at the target temperature by giving the target replica access to states which are only readily accessible at higher temperatures. The acceptance ratio for exchanges between replicas is governed by the overlap in the potential energy distributions associated with each replica. This leads to a main disadvantage of TREMD; that of requiring the replicas to be very closely spaced in temperature in order to generate a reasonable acceptance ratio. This requirement has the undesirable impact of demanding a large number of replicas to be used to span a given temperature range. The relationship between the temperature spacing between the replicas, and the number of replicas is well established and can be predicted [8].
To overcome the large number of required replicas inherent to many biological systems, Li et al. [9] developed the Temperature Interval with Global Energy Reassignment algorithm (TIGER) which they proceeded to further refine the algorithm resulting in the TIGER2 algorithm [10]. In both algorithms, the replicas are quenched to a baseline temperature and equilibrated at this baseline temperature before a single random exchange is attempted. This has the effect of reducing the number of accessible degrees of freedom between the replicas, increasing the acceptance rate. The other replicas are then re-arranged based on their potential energy. The reference implementation of Li et al. [9] was for the CHARMM MD package [11]. This algorithm was later ported to the NAMD [12] package as reported by Menz et al. [13] in which the replicas were controlled using a client–server paradigm implemented using the TCP/IP network stack between compute nodes. This method however is incompatible with certain high performance cluster architectures such as the BlueGene/Q (BG/Q) in which each node runs a highly optimised Compute Kernel, and inter-node communication is typically only possible via the PAMI/MPI libraries [14].
In this work we present a new implementation of the TIGER2 algorithm which makes use of the replica exchange Application Programming Interface (API) present in contemporary versions of NAMD, making it compatible across multiple HPC architectures including the BlueGene/Q and x86-based platforms. This work has been run on both the BG/Q and a Cray XC30, without the need to change any of the input files between the two platforms.
Section snippets
TIGER2 replica-exchange method
The TIGER2 method developed by Li et al. [10] attempts to overcome the limitations of the conventional T-REMD method by rather than attempting to exchange with the neighbouring replicas, by instead attempting the exchanges at the baseline temperature. Each TIGER2 simulation consists of concurrent simulations with each replica spaced between the baseline temperature () and the upper temperature limit (), and each given via: where takes values from 0 to . Each
Code description
The code takes the form of a TCL script which is called from within NAMD. The script makes use of the replica exchange API added into contemporary versions of NAMD (2.9). The script controls each replica and handles the exchanging of temperatures. The underlying exchanging of data and variables is handled by the MPI layer, or in the case of the BlueGene/Q the PAMI layer provided by IBM [14]. A pseudo-code description is presented in Algorithm 1.
Examples
The setup files for these examples are provided in the package containing the source-code, along with example job submission scripts suitable for both a BG/Q machine and a Cray XC30.
Summary
In this work we have presented an improved implementation of the TIGER2 REMD method, using the new REMD API found in contemporary versions of the NAMD Molecular Dynamics package. The implementation in this work improves on the previous implementation by bringing the TIGER2 method to the BG/Q platform in a flexible and consistent manner which allows for the use of the same input files between architectures. We have validated our implementation against two test systems, the first of which was
Acknowledgements
This work was supported by the Engineering and Physical Sciences Research Council [grant number EP/I001514/1]. This Programme Grant funds the Materials Interface with Biology (MIB) consortium. Computational time was provided by VLSCI, the Victorian Life Sciences Computation Initiative. TRW thanks veski for an Innovation Fellowship. AHB thanks Warwick University and Deakin University for studentship funding.
References (20)
- et al.
Replica-exchange molecular dynamics method for protein folding
Chem. Phys. Lett.
(1999) - et al.
Replica-exchange multicanonical algorithm and multicanonical replica-exchange method for simulating systems with rough energy landscape
Chem. Phys. Lett.
(2000) - et al.
TNAMD: Implementation of TIGER2 in NAMD
Comput. Phys. Commun.
(2010) - et al.
Replica exchange with solute scaling: a more efficient version of replica exchange with solute tempering (REST2)
J. Phys. Chem. B
(2011) - et al.
On easy implementation of a variant of the replica exchange with solute tempering in GROMACS
J. Comput. Chem.
(2011) - et al.
Efficient conformational sampling of peptides adsorbed onto inorganic surfaces: insights from a quartz binding peptide
Phys. Chem. Chem. Phys.
(2013) - et al.
Equation of state calculations by fast computing machines
J. Chem. Phys.
(1953) - et al.
Parallel tempering: Theory, applications, and new perspectives
Phys. Chem. Chem. Phys.
(2005) - et al.
A temperature predictor for parallel tempering simulations
Phys. Chem. Chem. Phys.
(2008) - et al.
An improved replica-exchange sampling method: temperature intervals with global energy reassignment
J. Chem. Phys.
(2007)
Cited by (0)
- ☆
This paper and its associated computer program are available via the Computer Physics Communication homepage on ScienceDirect (http://www.sciencedirect.com/science/journal/00104655).