How did you use the Xeon Phi co-processors of the Stampede supercomputer to investigate the complete influenza virus particle, or virion?
Historically, the influenza A virus has been responsible for millions of deaths worldwide each year. The persistence of seasonal H1N1 and H3N2 strains costs between 3,000 and 49,000 lives annually in the United States alone. Influenza pandemics, or global circulations of highly transmissible and pathogenic viral strains, have occurred four times in the past century. The 1918 H1N1 Spanish Flu was particularly devastating, killing as many as 100 million people worldwide.
We have built a full atomic model of the influenza envelope, comprised of 210 million atoms. In order to gain unprecedented insights into the mechanisms of influenza virulence that will accelerate anti-viral therapeutics development, we are running molecular dynamics (MD) simulations of this system on the Stampede Xeon Phi co-processors. The viral particle is the largest atomic-resolution system that has ever been simulated. Access to the Stampede Xeon Phi coprocessors has been critical for us because the size of the influenza system is too large to run on normal XSEDE CPU machines. Optimization of NAMD on the Stampede Xeon Phi's has enabled us to run simulations of the viral particle that otherwise would only be possible on the Leadership Class machines, which have limited resources available. Therefore, access to the Stampede Xeon Phi's has been critical to us being able to simulate this system.
We have used brief molecular dynamics simulations on Stampede to probe the stability of our model and to refine our new tools and technologies that others can use to build similarly complex or large systems. Our simulations have revealed certain regions of the model that require refinement before longer time scale simulations that will provide important, clinically relevant insights can be achieved.
In the process of creating this expansive model, we've developed a number of useful toolkits that will help other large-scale model builders as well. For example, we developed a python framework called pymolecule so computational biologists can more easily assemble multiple smaller molecular structures into models of expansive subcellular environments. We have also created and published a program called LipidWrapper that facilitates the construction of large-scale lipid-membrane bilayers of arbitrary geometry.
What contribution did XSEDE (through TACC) make to this research?
The large XSEDE allocation we received has been critical to our current research. Typical supercomputer allocations do not permit the simulation of atomistic models of this size. For example, Klaus Schulten and colleagues required a recent Blue Waters allocation to simulate the HIV-1 capsid, a system that is an order of magnitude smaller than our influenza particle. Without the help of resources like TACC and XSEDE, this work would simply not be possible.
How does this research impact both the scientific understanding and non-scientists?
These transformative simulations will enable an entirely new understanding of influenza structural biology – one in which the overall architecture of the influenza viral particle can be fully appreciated without loss of atomic-level details. Our work is an important first step on the path to fully understanding the influenza infection process. It will allow us to study the microscopic electric fields that surround these viral particles, fields that may impact virulence. It will also teach us more about the structure of key drug-binding sites on the viral surface, thus providing novel opportunities for drug and vaccine development.
What's the most important thing you want people to know about this research?
Influenza is itself a critically important virus to study, but the general potential of ever-larger atomic-resolution molecular models and simulations is even more exciting. There is a methodological gap in structural biology that can only be answered with these kinds of projects. Techniques like X-ray crystallography and NMR are capable of producing atomic-resolution models of individual macromolecules (e.g., proteins), and electron microscopy is capable of producing lower-resolution models of larger multi-component systems. However, no current experimental technique can consistently combine the advantages of these methods to produce atomic-resolution models of whole large-scale, multi-component subcellular environments. In silico modeling and simulation together constitute a powerful "computational microscope" that bridges these experimental techniques. It is only in recent years that simulations on this scale have become possible.