Christopher Lockhart
Christopher Lockhart
Ph.D. Candidate
George Mason University

28 November 2014 — by Chris Lockhart

One of the benefits of the Compute Against Alzheimer's Disease campaign is the introduction of novel software to perform large-scale molecular dynamics simulations. The specific method used in this software is called replica exchange molecular dynamics (REMD). Usually, a high-powered computer cluster as can be found through Amazon or XSEDE is required to run REMD simulations. Our novel software alleviates this requirement and allows these simulations to be performed remotely, on personal computers scattered around the globe.

The difference in performance between REMD and conventional molecular dynamics (CMD) is easy to imagine. CMD generally runs a system (such as the Aβ peptide in water) at a fixed temperature and volume (or pressure) for a certain amount of time. The molecular processes of binding and structural transitions which we are interested in studying take 100 nanoseconds to years to complete naturally in biological systems, such as the cells in your body. CMD simulations typically run with a time step of 1 femtosecond. Therefore, each time evolution of a CMD simulation only gets 10-6% closer to the goal of about 100 nanoseconds. The amount of wall clock time required to complete 100 nanoseconds varies by system size. Using benchmarking done by the University of Illinois at Urbana-Champaign, it takes roughly a week using 4 CPU cores to simulate just a single nanosecond. It would thus take in excess of 100 weeks of computation to get in the range of timescales of molecular motions that we want to observe.

REMD circumvents many of the shortcomings seen above by slightly changing the strategy. Rather than simulating a system at a fixed temperature, REMD simulates a system at a range of temperatures at once. As the simulations progress, the simulations from higher temperatures are used to help the simulations at lower temperatures adopt new conformations, which has the end effect of drastically reducing the timescale of molecular motions of interest. A direct comparison of CMD and REMD is presented in Figure 1, where in the same amount of simulation time REMD is able to generate approximately three times the amount of protein conformations as CMD. The simulation time curve for CMD in Figure 1 can be fit with a logarithmic function, which allows us to predict that CMD would have to run for 2.7 seconds (7 orders of magnitude more) of simulation time to generate the same amount of protein conformations as REMD. Clearly use of CMD is unreasonable, and REMD presents a much more efficient method of producing results.

At present, the Compute Against Alzheimer's Disease campaign is the only distributed computing campaign actively using REMD. We are confident that with this sophisticated method we will be able to generate reliable and important results relevant to the study of Alzheimer's disease.

Figure 1. Plot of the number of protein conformations sampled by replica exchange molecular dynamics (black) and conventional molecular dynamics (red) as a function of simulation time. Simulations of Aβ16-22 in water were used to generate this data.