CUDA Spotlight: Aron Broom

Aron Broom

GPU-Accelerated Molecular Dynamics

This week's Spotlight is on Aron Broom. Aron is a researcher at the University of Waterloo in Canada, where he works on the molecular engineering of protein structure and function.

Aron's current research interests include protein-ligand binding - particularly of a multivalent nature - and stability and symmetry in protein structure and evolution.

This interview is part of the CUDA Spotlight Series.

Q & A with Aron Broom

NVIDIA: Aron, how did you become involved in Molecular Dynamics (MD)?

Aron: I started working on protein receptor site design (the "lock") and ligand drug candidates (the "key") binding largely from an experimental perspective, but without extremely labor and resource intensive techniques, one often doesn't get a sense of what is happening at a detailed molecular level.

For instance, you can see from a "dry" crystal structure where on a protein a particular compound binds, but since these bindings occur in liquid water you don't know what kind of movements were needed to get it there, hold it there via non-bonding forces or which parts are most critical for that interaction (it's not always the parts closest to one another).

I wanted to know what was happening at a molecular level. I wanted to "see" what the atoms were doing. MD was the perfect tool and GPUs make the simulations much more affordable and accessible to the "common man."

NVIDIA: Give us a quick description of MD.
Aron: MD offers an opportunity to study biological interactions at the atomic level, by running a simulation of the particles making up the molecules of interest under the effect of the forces that govern their interactions.

For example, MD allows scientists to study how tightly a drug compound will bind a medically relevant protein or other target (carbohydrate, nucleic acid, or lipid). But more importantly, the high level of detail allows us to understand and subsequently redesign a better drug candidate for that existing binding site.

MD simulations often involve thousands to millions of atoms, all interacting with one another through several different forces. But simply computing these interactions once only gives a single snapshot. In order to understand the dynamics and properties of the molecules of interest we must repeat these large calculations millions, billions or even trillions of times; GPUs have been most helpful in that we can now simulate larger systems for the longer timeframes needed to see the final equilibrium state obtained.

Molecular Dynamics in Explicit Solvent: Protein-ligand interactions can be studied through MD by solvating those molecules in a box of hundreds to thousands of water molecules. The properties of the molecular system, such as the binding affinity of the protein for that ligand can then be calculated.

NVIDIA: What are some key challenges in your field?
Aron: An important challenge facing modern molecular biology is not only to understand the molecular interactions between the constituents of cells (proteins, nucleic acids, lipids, and carbohydrates) in order to predict their behavior, but ultimately, to redesign and engineer them ourselves in order to control existing functions and design new ones. It's a challenge that I'm particularly interested in helping to solve.

Furthermore, in order to accurately calculate the properties I'm interested in, not only does each individual simulation need to be long enough to capture the relevant configurations of the molecules involved, but it is often necessary to run multiple simulations, each of which examines the behavior of the molecules under different conditions and to statistically analyze these multiple independent "movies" to see what is most likely to occur in nature.

Overall, this creates a situation in which the computational demand is extremely high, and solving the problem using CPUs would simply not be feasible under realistic conditions where large high-powered computing resources need to be shared with many other scientists. Because of the incredible performance to price and performance to power ratios of GPUs, I'm able to study problems using small GPU clusters which would not be possible using larger traditional CPU-only clusters.

NVIDIA: What are some examples of your focus areas?
Aron: Using GPUs I've been able to study protein-ligand interactions across a number of different molecular systems:

NVIDIA: When does it make sense to use a single GPU workstation rather than a cluster?
Aron: While Tesla GPU-based HPC clusters have incredible potential for allowing the volume of computations needed to obtain accurate results, a single consumer GPU has its place in allowing rapid experimentation and development without needing to wait for queue-ing times on HPC clusters. In particular I've found OpenMM, an MD library designed specifically for GPUs, extremely useful as I'm able to quickly code my own custom simulations which have absolutely astonishing performance.

For instance, implicit solvent simulations (where water is treated as a continuum rather than individual atoms) benefit to an extraordinary extent from running on a GPU. Using a single GeForce GTX570 in my desktop, I'm able to simulate protein folding and protein-protein interactions of a moderately-sized molecular system at a speed that would require ~400 CPU cores if running in NAMD (which is known for nearly linear scaling across multiple CPU cores) on a modern HPC cluster.

Given that the GTX570 itself has 480 cores, we can see that for this application a GPU and CPU compare almost core-to-core, and given that newer GPUs contain thousands of cores, the possibilities for the future are staggering.

Molecular Dynamics in Implicit Solvent: When water molecules are included explicitly they create a viscosity, which is realistic, but can slow exploration of molecular movements such as protein folding. Implicit solvent offers an attractive alternative, representing the water as an electrostatic-modulating medium, which may capture many of the bulk effects while allowing unhindered movements of the protein.

NVIDIA: What's next on the horizon?
Aron: In my opinion, the primary drawback to using GPUs is having a critical mass of applications and features designed to benefit from GPU performance enhancement. Fortunately, the incredible benefits GPUs offer is driving progress in this regard.

For instance the following molecular dynamics programs have had many or all of their components ported to a GPU version: AMBER, NAMD, LAMMPS and GROMACS. And, OpenMM is designed expressly for use on a GPU.

Given the incredible explosion in GPU performance, with the number of cores more than quadrupling over the last year alone, I expect we're quite rightfully seeing a shift to a new standard where GPU enhancement becomes the norm. Overall it's a thrilling time to be doing computational science!

Bio for Aron Broom

Starting as an experimentalist, Aron spent much of his time running around the lab with vials of purified protein. After completing a Master's in biochemistry he switched to a computational perspective and has been sitting at a computer ever since. In his ever-decreasing spare time he enjoys canoeing, rock-climbing, and any coding or electronics tinkering he can manage.

Relevant Links
Explicit ligand binding video
Implicit unfolding video


Compute Canada

Contact Info
rabroom (at) uwaterloo (dot) ca
Earth Sciences and Chemistry Building, Room 230A,
University of Waterloo, Waterloo, ON, Canada, N2L 3G1