For centuries, influenza epidemics have been a major cause of mortality worldwide, with large numbers of people affected every year. To this day, widespread research continues around the world as scientists strive to learn more about the behavior and structure of the influenza virion and its life cycle, with the goal of reducing or eliminating influenza infections and epidemics through safer and more effective antiinfluenza drugs.
Historically, studying viruses such as influenza in laboratory experiments has been difficult because reactions produce intermediate products that are temporary and too unstable to capture. In addition, previous attempts at simulating these systems were beyond the reach of supercomputers, due to the complexity of simulating billions of particles under the right environmental conditions.
At the Institute of Process Engineering, Chinese Academy of Sciences (CAS-IPE) in Beijing, researchers have used a GPU-based heterogeneous supercomputer to create the world’s first simulation of a whole H1N1 influenza virus at the atomic level.
With this new level of visibility, scientists can help bridge the gap between biology, virology, epidemiology, and drug development at the molecular level, potentially leading to new and more powerful drug treatments and vaccines to combat influenza.
Current study of influenza virus
Today, there is a huge gap between the way research scientists study viruses and the way pharmaceutical developers create antiviral drugs and vaccines. Virologists provide a general picture of virion particles by examining structural changes during the virus’ life cycle (e.g., binding to the cell, uncoating the viral particle, replicating itself using the genetic material of the cell, assembly and release from the host cell). Influenza virions are highly polymorphic, with sizes ranging from spherical particles with a diameter of approximately 100 to 150 nm, to filamentous particle with a length of several millimeters.1 The surface of the influenza virion is characterized by distinctive spikes, HA (hemagglutinin) and NA (neuraminidase), with an approximate ratio of four HA to one NA.
However, research at this level provides only limited information. It does not provide deep insight into the internal chemical structure or biological behavior of the influenza virion, which is required to help pharmaceutical companies develop more effective vaccines and drugs.
On the molecular level, biologists resolve the structure of key proteins of the virion particle and find potential targets within these proteins, which aids pharmaceutical developers who design antiinfluenza drugs and vaccines. The 3D structures of all the trans-membrane proteins, such as NA, HA, and M2, have already been resolved, though some other protein structures are not yet complete. Eight ribonucleoprotein (RNP) complexes exhibit regular helical conformations respectively, and are located in the interior of the virion.2
Drug developers analyze the properties of the trans-membrane proteins, find their potential targets and design antiinfluenza drugs based on these intended targets. Some drugs, like zanamivir and oseltamivir, are designed to inhibit NA activity, while other antiviral drugs are designed according to the structure of other proteins within a virus. Adamantanes, for example, function by blocking the M2 channel.
Though both research scientists focused on basic biology and pharmacologists focused on clinical development of drugs welcome the support from each other, it is still difficult to bridge their studies since the experimental facilities and studies are expensive and have limited resolving power of time, space and environment (the condition where the virion lives: temperature, PH, ionic concentration, etc.).
Constructing a model
Previously, molecular dynamic (MD) simulation has been used as a “computational microscope” to probe the atomic structure of biological molecules and detect dynamic processes on small spatio-temporal scales. To start the simulation, the first step is to construct a static molecular model of the complete virion particle. The chemical, structural, and biological information have already provided a stationary picture of the 3D structure of influenza virions.The model, however, is still rough and the structures of some component molecules are still unknown, or information on them is incomplete. These missing details should be reconstructed to provide an atomic structure of the virion, and then an explicit solvent MD study should be performed to further detect the dynamics of the virion in vivo.
Based on the model of influenza vRNPs, the entire atomic structure of a single nucleoprotein molecule is first constructed using the crystal structure, and second, nucleoprotein monomers are placed in a helical structure. On the surface of each protein scaffold, a packed single-stranded negative-sense RNA strand with 924 to 2,377 nucleotides3 is tightly located as helical structures. All eight vRNPs are closely spaced to form the interior of the virion. The spherical protein layer of M1, which coats the vRNPs, is constructed using many copies of the crystal structure of a single M1 molecule4 located in a sphere. Since dipalisitoylphosphatidylcholine (DPPC) is a main component of the membrane, the DPPCs are evenly spaced on a spherical surface with double layers, resulting in a globular membrane with an external diameter of 106 nm.
Three trans-membrane proteins—NA, HA, and M2—are constructed following similar procedures. Specifically, a single macromolecular structure is reconstructed at the atomic-scale, followed by preparation of a matrix of the macromolecules on a spherical surface. Only the ectodomains of NA and HA have been resolved by X-ray crystallographic techniques, corresponding to residues 83 to 468 of NA5 and residues 11 to 325 of HA.6 The endodomains, however, are difficult to crystallize since they are insoluble. According to protein structure prediction,7 residues 11-31 of NA and residues 14-40 of HA should be the trans-membrane segment and be helical in shape, while the other endodomain residues tend to be random coils. The 3D structures of the endodomain are thus constructed using PyMol (Schrödinger) and Visual Molecular Dynamics.8
After a short time for dynamic simulation, a comparative stable structure of the protein can be obtained. 374 HA and 98 NA molecules are located on a sphere with their tails embedded in the lipid membrane as appropriate. M2 is a single-pass membrane protein, and the proton channel is formed by four parallel monomers. The structure of a whole proton channel is first constructed based on NMR results,4 and then many copies are located on the spherical surface with the trans-membrane segments embedded in lipids. After deletion of overlapped molecules, the simulated H1N1 virion appears constructed with 2,363 proteins, 63,471 DPPC molecules, and 8 RNA strands (Figure 1).
Molecular simulation of H1N1
After construction of the molecular model with atomic details, the entire complex is solvated in water with an appropriate concentration of ions to represent the environment in vivo. The system has 300 million atoms locating in a periodic cube with each side measuring 148.5 nm long.
Traditional CPU-based MD software and hardware are incapable of simulating the dynamics of such a large biological system, as the large number of CPU nodes required for timely results would be costly and space prohibitive. However, researchers have been able to dramatically increase computing power for molecular dynamics and other scientific applications by using efficient, high-performance graphics processors (GPUs) that serve as companion processors to the CPU.
These high-performance hybrid supercomputing systems not only allow researchers to run complex scientific applications and simulations significantly faster than on CPU-only systems, they also enable the simulation of large, realistic biological systems that had previously not been possible. To reduce computational load, earlier viral simulations generally used a coarse-grained method, treating tens or even hundreds of atoms as one bead. However, the empirical parameterization of the beads meant that the simulation results were less reliable. In addition to giving scientists far greater visibility into the molecular structures and biological behaviors of the virus, simulations can now be run in hours or days, rather than weeks or months.
To simulate the H1N1 virus on the atomic level, scientists at CAS-IPE used the GPU-based supercomputer, Mole-8.5 (Figure 2), which enabled them to observe the dynamic structure of H1N1 virion.
The Mole-8.5 supercomputer can deliver a peak performance of over 1 petaflops, which places it 21st on the TOP500 list of the world’s most powerful supercomputers. It is also ranked ninth on the annual Green500 list, which tracks the world’s most energy-efficient supercomputers.
With a custom MD software package,9 CAS-IPE scientists ran the influenza simulation on 288 low-level hybrid computing nodes consisting of 1,728 NVIDIA Tesla C2050 GPUs, which reached a speed of 0.77 ns/day with an integration time step of 1 femtoseconds. Starting from the predefined structure, the virus experiences significant shape and energy changes over the timescale simulated until one obtains a stable energy minimized conformation. The structural and energetic changes of each component can then be analyzed from the dynamic calculations.
Using this model, researchers can more easily experiment with a variety of protein targets and drug candidates under different environments and conditions, and observe in great detail how potential treatments interact with the influenza virion.
In addition, potential targets can be identified through an analysis of the atoms and portions of the protein molecules that play key roles in the life cycle of the virus. At the same time, new drugs can be designed to bind more effectively and efficiently to the targets, resulting in increased efficacy, safety, and a reduced life cycle for the virus. Starting from simulating a certain number of drug candidates in solution, researchers can look at the binding process of drugs to the potential targets, and the succeeding behavior of the virion particle to prioritize which drug candidates may be safest and most effective in vivo. They can also use this model to investigate the response of the virion to an external mechanical force, e.g., inhalation or extrusion.
All these studies require longer simulation times, from tens to hundreds of nanoseconds, or longer. With the added performance of GPU-based hybrid supercomputers and further optimization of scientific algorithms, virus simulations could run at a higher speed even while evaluating larger systems of interest. Alternatively, researchers can simulate in a coarse-grain fashion for some sub-parts of the virion that are assumed to be non-pivotal, while analyzing the most important molecules within the virion more rigorously at atomic scale.
This type of approach to the simulation of large, complex biological systems holds great promise for science, potentially enabling a wave of new breakthroughs in the ability to understand and battle infectious disease.
About the Authors
Ji Xu is interested in algorithm design of molecular dynamics simulation and other particle simulation methods like discrete element method. He was mainly responsible for developing the MD codes for simulating the virion. Ying Ren’s research focuses on molecular dynamics simulations of bio-systems. She was responsible for construction of the virion structure based on current experimental and theoretical understandings. Wei Ge studied multi-scale discrete simulation of complex systems. He served as a coordinator for the software and hardware development involved in this study. Xiaowei Wang focused on GPU cluster hardware development and large-scale discrete simulations based on GPU. Xianfeng He studied parallel visualization and GPU application in process engineering.
1. Roberts PC, Compans RW. Host cell dependence of viral morphology. PNAS.1998;95(10):5746-5751.
2. Wu WWH, et al. Ultra-structural Analysis of the Nuclear Localization Sequenceson Influenza A Ribonucleoprotein Complexes. J Mol Bio. 2007;374(4):910-916.
3. GenBank. Available from http://www.ncbi.nlm.nih.gov/genbank. Accessed on March 7, 2012.
4. Stouffer AL, et al. Structural basis for the function and inhibition of an influenza virus proton channel. Nature. 2008;451(7178):596-599.
5. Russell RJ, et al. The structure of H5N1 avian influenza neuraminidase suggests new opportunities for drug design. Nature. 2006;443(7107):45-49.
6. Xu R, et al. Structural Basis of Preexisting Immunity to the 2009 H1N1 Pandemic Influenza Virus. Science. 2010; 328(5976):357-360.
7. The PSIPRED Protein Structure Prediction Server. University College London. Available from http://bioinf.cs.ucl.ac.uk/psipred. Accessed on March 7, 2012.
8. Humphrey W, Dalke A, Schulten K. VMD: Visual molecular dynamics. J Molecular Graphics. 1996;14(1):33-38.
9. GPU-based MD programs. EMMS Group. Available from http://emms.mpcs.cn/software/810.htm. Accessed on March 7, 2012.