Research Highlights

NBCR Researchers Publish New Molecular Dynamics Method that Enhances Biomolecular Sampling by Orders of Magnitude and Works with a Polarizable Force Field

box

Figure 1: First author Steffen Lindert in his office at UCSD.

A team of NBCR researchers recently demonstrated a new accelerated molecular dynamics (aMD) implementation that enhances sampling of the conformational space of biomolecules by several orders of magnitude. More specifically, they developed an implementation of the aMD method using the OpenMM toolkit library to produce high performance – and, as a result, longer simulation times – on graphics processing units (GPUs). Furthermore, they showed that this method works with AMOEBA (in a joint AMOEBA-aMD implementation) to simulate relatively long-time-scale events with a polarizable force field – both efficiently and with accurate results.

They have made their work publicly available (http://wiki.simtk. org/openmm/VirtualRepository) and published their results in the October 15, 2013, issue of the Journal of Chemical Theory and Computation (see citation below).

The research team consisted of Steffen Lindert, Denis Bucher, Peter Eastman, Vijay Pande, and J. Andrew McCammon. Together, they represent the departments of Pharmacology, Bioengineering, and Chemistry & Biochemistry, the National Biomedical Computation Resource, the Center for Theoretical Biological Physics and the Howard Hughes Medical Institute, all at UCSD, and Simbios, the NIH Center for Biomedical Computation at Stanford University. They ran their simulations on hardware supported and managed by NBCR through a GPU allocation to McCammon.

Says McCammon, “Having served for many years on the Scientific Advisory Board of Simbios, which is one of NBCR’s sister organizations, I’m pleased to be part of this collaboration between the NIH resources at Stanford and UCSD. Steffen is an outstandingly adventurous and talented postdoc. His willingness to spend time at Stanford really powered this project.”

Molecular dynamics (MD) is a well-established technique to study the dynamics and equilibrium properties of biomolecules. It has progressed over the last 35 years to enable simulations on the order of microseconds of increasingly larger macromolecular systems. But many important biological processes require millisecond or longer simulations. Typically, researchers who had access to NSF and other supercomputers moved their simulations there. But the relatively recent emergence of GPUs has help address this constraint for many by providing broader and cheaper access to higher computational power. Even so, researchers still lack sufficient power to capture the Holy Grail: the ability to characterize the free-energy landscapes of the largest molecular systems.

The fallback solution, for now, has been enhanced sampling methods. Among those available, aMD is a promising technique to modify regions of the potential energy landscape below a certain energy cutoff. It has been implemented, for example, in classical (nonpolarizable) simulations in AMBER and NAMD.

But improved sampling is only part of the problem researchers face. In addition, they need to improve the force fields that propagate the atoms. These are related problems, as improved sampling enables better detection of force field inaccuracies. As a result, researchers have focused on developing force fields with more sophisticated functional forms, including polarizable force fields.

In this context, the team turned to AMOEBA (Atomic Multipole Optimized Energetics for Biomolecular Applications), one of the most commonly used polarizable force fields. But even with GPU acceleration, AMOEBA use has been limited, since its simulations of large biomolecular systems are 1-2 orders of magnitude more computation intensive than those of their nonpolarizable counterparts.

But the team saw an opportunity: Combine AMOEBA with aMD by way of OpenMM, a molecular dynamics toolkit that has attracted a wide following because it leverages the new GPUs to accelerate simulations.


Figure 2: Integration of AMOEBA with aMD through OpenMM

Lindert brings the process down to earth: “We were studying a protein that binds calcium, which is a highly charged ion. Many of the properties, such as binding affinities, though, didn’t calculate correctly with nonpolarizable methods. As a result, we were pushed to use polarizable force fields. In that regard, AMOEBA is state of the art, but it’s not as fast as we needed, so we had to integrate it with aMD to compensate. aMD had not yet been implemented with OpenMM, which was the ticket to GPUs and better performance. That motivated our collaboration with Simbios, the experts in both AMOEBA and OpenMM.”

Lindert applied to Simbios’ visiting scholar fellowship program with a proposal to implement aMD with OpenMM. He was accepted and spent a month at Stanford doing the work. He emphasizes the importance, even in the age of the Internet, of face-to-face contact to propel research: “It was extremely helpful to be on site at Stanford where I had contact with colleagues over lunch every day to talk through problems. My experience has been that once you know someone personally and understand how he thinks, e-mail can become a pretty good substitute. But you need to establish the relationship in person first.”

After completing the integration work, the team ran three types of aMD – one version to boost the dihedral potential, a second to boost the total potential, and a third to boost both – using OpenMM’s CustomIntegrator (http://wiki.simtk.org/openmm/VirtualRepository) that enables implementing customized user algorithms. (They also ran conventional molecular dynamics [cMD] simulations for comparison.)

The team, more specifically, studied three systems:

  • Alanine dipeptide, which they describe as the “workhorse of molecular simulations,” to make sure the aMD implementation was done correctly and did indeed improve sampling. By comparison with cMD, the dihedral-boost and dual-boost implementations sampled the conformational space much more efficiently.
  • Bovine pancreatic trypsin inhibitor (BPTI), to show that the combined AMOEBA-aMD simulations were stable even at high acceleration levels. The team achieved an acceleration of sampling by 3-4 orders of magnitude.
  • The endonuclease IV metallo-enzyme, because it’s a large protein with a stable three-zinc cluster that has not been well described in nonpolarizable simulations. By comparison with nonpolarizable AMBER simulations, their polarizable simulations more accurately described the geometry of the cluster.

These researchers believe that their aMD sampling method as implemented in OpenMM should effectively support studies that require more extensive (micro- to millisecond) biomolecular sampling. Their results also show preserved accuracy of the AMOEBA polarizable force field with improved sampling efficiency (typically by 2-3 orders of magnitude). Finally, AMOEBA better described the active site of the endonuclease zinc cluster, reflecting the results of previous studies.

References: Lindert, Steffen, Denis Bucher, Peter Eastman, Vijay Pande, and J. Andrew McCammon, Accelerated Molecular Dynamics Simulations with the AMOEBA Polarizable Force Field on Graphics Processing Units, Journal of Chemical Theory and Computation, 2013, 9, 4684-4691, dx.doi.org/10.1021/ct400514p.

NBCR Researchers:Lindert, Steffen, Denis Bucher, Peter Eastman, Vijay Pande, and J. Andrew McCammon

Figure 1:First author Steffen Lindert in his office at UCSD.

Figure 2:Integration of AMOEBA with aMD through OpenMM.