md simulations with namd (and vmd)
TRANSCRIPT
PRACE/BioExcel Spring School 2019HPC for Life Sciences
Sweden
MD Simulations with NAMD (and VMD)
João V. Ribeiro
www.ks.uiuc.edu/[email protected]
University of Illinois at Urbana-Champaign NIH Center for Macromolecular Modeling and Bioinformatics
Research Programmer
A Brief History (and Future) of NAMD and VMDNumber of Citations per Year
VMD NAMD
Hours Until Next Citation (VMD + NAMD)
NAMD in a Nutshell• Developed in C++, CUDA (GPU), and Charm++
• Performance Scales to Hundreds of Thousands of
Cores and Hundreds of GPUs
- Large Systems
- Enhanced Sampling
• Large Variety of User Defined Forces and Biased
Simulations
• TCL Script as Input File
- Allows Scripting in the Input File
- Workflow Control
- Method Development at Higher Level
• Close Relationship with VMD
- Preparation - QwikMD
- Analysis - Cross Correlation, Clustering…
- Visualization- Ray Tracing
IEEE Fernbach Award 2012 - “For outstanding contributions to the development of widely used parallel software for large biomolecular systems simulation”
NAMD: http://www.ks.uiuc.edu/Research/namd/VMD: https://www.ks.uiuc.edu/Research/vmd/
E.Coli Chemosensory Array Protocell
Main NAMD Developers and Contributors
Jim Philips NCSA Blue Waters
David Hardy Senior Research Programmer
Julio Maia Research Programmer
Brian Radak Research Programmer
Giacomo FiorinTemple University
Jérôme HéninInstitut de Biologie
Physico-Chimique (Paris)
Ryan McGreevy Research Programmer
Wei Jiang Argonne Lab
NAMD Developer WorkshopI NAMD Developer Workshop - Chicago, IL 2016 II NAMD Developer Workshop - Chicago, IL 2017
III NAMD Developer Workshop - Urbana, IL 2018
NAMD Developer WorkshopI NAMD Developer Workshop - Chicago, IL 2016 II NAMD Developer Workshop - Chicago, IL 2017
III NAMD Developer Workshop - Urbana, IL 2018
Upcoming NAMD Developer Workshop August 19-20 2019
Urbana Illinois
http://www.ks.uiuc.edu/Training/Workshop/Urbana2019/
Hands-On NAMD• 55 Workshops • 50+ Tutorials
- 5 New Tutorials- 1800+ pages of Tutorials
• 12 Case Studies• Hands-On Workshop on Enhanced Sampling and
Free-Energy Calculation - September (2019) - to be Announced
Hands-On Workshops
Training - https://www.ks.uiuc.edu/Training/Previous Workshop Streams - https://www.youtube.com/user/tcbguiuc/playlists
NAMD 2.13 - What’s New
NAMD: https://www.ks.uiuc.edu/Research/namd/2.13/features.html
• Stochastic Velocity Rescaling Thermostat
• Replica Exchange with Solute Scaling (REST2)
• Hybrid Quantum Mechanics/Molecular Mechanics (QM/MM) Simulation
• Interleaved Double-Wide Sampling for Alchemical FEP
• Constant-pH MD
• Gaussian accelerated MD (GaMD)
• τ-Random acceleration MD (τRAMD)
• Improved Support for Lone pair and Polarizable Drude Force Field
• Scaling on Summit Supercomputer
• Support for billion-atom systems
NAMD 2.13 - What’s New
NAMD: https://www.ks.uiuc.edu/Research/namd/2.13/features.htmlNAMD On Summit: http://www.ks.uiuc.edu/Research/namd/2.13/NAMD-IBM-Journal-Manuscript-Revised.pdf
• Stochastic Velocity Rescaling Thermostat
• Replica Exchange with Solute Scaling (REST2)
• Hybrid Quantum Mechanics/Molecular Mechanics (QM/MM) Simulation
• Interleaved Double-Wide Sampling for Alchemical FEP
• Constant-pH MD
• Gaussian accelerated MD (GaMD)
• τ-Random acceleration MD (τRAMD)
• Improved Support for Lone pair and Polarizable Drude Force Field
• Scaling on Summit Supercomputer
• Support for billion-atom systems NAMD pre-2.13, STMV matrix, 2fs timesteps 5x2x2 STMV ≈ 21M atoms
7x6x5 STMV ≈ 224M atoms
NAMD 2.13 GPU Performance Improvements
• Simulation Parameters:- Integration Time Step: 1 fs- Cutoff: 12- Switch: 10- CHARMM Force Field
• Ivy Bridge system: dual socket Intel Xeon CPU E5-2690 v2 @ 3.00 GHz, 20 total cores.
• Haswell system: dual socket Intel Xeon CPU E5-2698 v3 @ 2.30 GHz, 32 total cores
• Skylake system: dual socket Intel Xeon Gold 6148 CPU @ 2.4 GHz, 40 total cores.
Apolipoprotein A-I NAMD: http://www.ks.uiuc.edu/Research/namd/benchmarks/
NAMD 2.13 GPU Performance Improvements
Satellite Tobacco Mosaic Virus
• Simulation Parameters:- Integration Time Step: 1 fs- Cutoff: 12- Switch: 10- CHARMM Force Field
• Ivy Bridge system: dual socket Intel Xeon CPU E5-2690 v2 @ 3.00 GHz, 20 total cores.
• Haswell system: dual socket Intel Xeon CPU E5-2698 v3 @ 2.30 GHz, 32 total cores
• Skylake system: dual socket Intel Xeon Gold 6148 CPU @ 2.4 GHz, 40 total cores.
NAMD: http://www.ks.uiuc.edu/Research/namd/benchmarks/
Exemplary NAMD Features• User Defined Forces
- Grid Forces- Interactive Molecular Dynamics- Steered Molecular Dynamics
• Accelerated Sampling Methods- Replica Exchange
• Collective Variable (Colvars)
- Biased Simulations
- Enhanced Sampling
• Free-Energy Calculation Methods
- Free-Energy Perturbation
- Adaptative Biasing Force
- Constant pH Simulations . . .• Hybrid QM/MM Simulations
Complete List of NAMD Features: https://www.ks.uiuc.edu/Research/namd/2.13/ug/
Grid Forces
Trabuco et al. Structure (2008)Trabuco et al. Methods (2009)Wehmer et al. PNAS (2017)Grid Forces - https://www.ks.uiuc.edu/Research/namd/2.13/ug/node41.htmlMDFF- https://www.ks.uiuc.edu/Research/mdff/MDFF Tutorial - https://www.ks.uiuc.edu/Training/Tutorials/#mdff
• Addition Potential Term• Arbitrary Shape and Magnitude• Three Dimensional Grid with Scaling
factor on Each Voxel
• Use VMD to “Translate” Density Data into Potential Grid
UEM(R) = ∑j
wjVEM(rj),
Grid Forces for Molecular Dynamics Flexible Fitting
• Addition Potential Term• Arbitrary Shape and Magnitude• Three Dimensional Grid with Scaling
factor on Each Voxel
• Use VMD to “Translate” Density Data into Potential Grid
- Molecular Dynamics Flexible Fitting
UEM(R) = ∑j
wjVEM(rj),
VEM(r) =ξ(1 −
Φ(r) − Φthr
Φmax − Φthr ), if Φ(r) ≥ Φthr
ξ, if Φ(r) < ΦthrTrabuco et al. Structure (2008)Trabuco et al. Methods (2009)Wehmer et al. PNAS (2017)Grid Forces - https://www.ks.uiuc.edu/Research/namd/2.13/ug/node41.htmlMDFF- https://www.ks.uiuc.edu/Research/mdff/MDFF Tutorial - https://www.ks.uiuc.edu/Training/Tutorials/#mdff
Molecular Dynamics Flexible Fitting
Supercomputer
Electron Microscope
EM densitymap
crystallographic structure
Match through MD
(Ribosome-bound YidC)
APS Synchrotron
Trabuco et al. Structure (2008)Trabuco et al. Methods (2009)Wehmer et al. PNAS (2017)Cassidy et al. eLife (2015)Chemotaxis http://www.ks.uiuc.edu/Research/chemotaxis/MDFF- https://www.ks.uiuc.edu/Research/mdff/MDFF Tutorial - https://www.ks.uiuc.edu/Training/Tutorials/#mdff
Molecular Dynamics Flexible Fitting (MDFF)
Integrating experimental data to produce models of biomolecular complexes with atomic detail
E.Coli Chemosensory Array
Proteasome
Molecular Dynamics Flexible Fitting (MDFF)
Integrating experimental data to produce models of biomolecular complexes with atomic detail
E.Coli Chemosensory Array
Proteasome
Trabuco et al. Structure (2008)Trabuco et al. Methods (2009)Wehmer et al. PNAS (2017)Cassidy et al. eLife (2015)Chemotaxis http://www.ks.uiuc.edu/Research/chemotaxis/MDFF- https://www.ks.uiuc.edu/Research/mdff/MDFF Tutorial - https://www.ks.uiuc.edu/Training/Tutorials/#mdff
Molecular Dynamics Flexible Fitting (MDFF)
Integrating experimental data to produce models of biomolecular complexes with atomic detail
Cascade MDFFHigh Resolution Density Maps
E.Coli Chemosensory Array
Proteasome
Trabuco et al. Structure (2008)Trabuco et al. Methods (2009)Wehmer et al. PNAS (2017)Cassidy et al. eLife (2015)Chemotaxis http://www.ks.uiuc.edu/Research/chemotaxis/MDFF- https://www.ks.uiuc.edu/Research/mdff/MDFF Tutorial - https://www.ks.uiuc.edu/Training/Tutorials/#mdff
Interactive Modeling with MDFF GUI
Set up and run interactive (or traditional) MDFF/xMDFF simulations
Analyze interactive simulations in real-time
• Apply forces to manually manipulate structure into the density• Useful for difficult to fit structures with large conformational changes
Modeling Large Complex Membrane Systems
Distribution of proteins across the membrane surface (dense environment)• Ability the handle a variety of protein geometries• Proper orientation of proteins in relation to the
membrane surface• Generalizable and automated method for
membranes of arbitrary shapeEmbedding proteins into the membrane• Account for surface area occupied by proteins in
inner and outer leaflets• Proper lipid packing around embedded proteins
Vesicle Construction Coarse Grain Protein CG Protein Placement Combine Lipid + Protein
Vesicle Construction Coarse Grain Protein CG Protein Placement Combine Lipid + Protein
Modeling Large Complex Membrane Systems
Distribution of proteins across the membrane surface (dense environment)• Ability the handle a variety of protein geometries• Proper orientation of proteins in relation to the
membrane surface• Generalizable and automated method for
membranes of arbitrary shapeEmbedding proteins into the membrane• Account for surface area occupied by proteins in
inner and outer leaflets• Proper lipid packing around embedded proteins
Vesicle Construction Coarse Grain Protein CG Protein Placement Combine Lipid + Protein
Modeling Large Complex Membrane Systems
Distribution of proteins across the membrane surface (dense environment)• Ability the handle a variety of protein geometries• Proper orientation of proteins in relation to the
membrane surface• Generalizable and automated method for
membranes of arbitrary shapeEmbedding proteins into the membrane• Account for surface area occupied by proteins in
inner and outer leaflets• Proper lipid packing around embedded proteins
Vesicle Construction Coarse Grain Protein CG Protein Placement Combine Lipid + Protein
Modeling Large Complex Membrane Systems
Distribution of proteins across the membrane surface (dense environment)• Ability the handle a variety of protein geometries• Proper orientation of proteins in relation to the
membrane surface• Generalizable and automated method for
membranes of arbitrary shapeEmbedding proteins into the membrane• Account for surface area occupied by proteins in
inner and outer leaflets• Proper lipid packing around embedded proteins
0.4 μm
113 million Martini particlesrepresenting 1 billion atoms
3.7 M lipids (DPPC), 2.4 M Na+ & Cl- ions, 104 M water particles (4 H2O / particle)
Protein ComponentsAquaporin Z Copper Transporter (CopA) F1 ATPase Lipid Flipase (MsbA) Molybdenum transporter (ModBC) Translocon (SecY) Methionine transporter (MetNI) Membrane chaperon (YidC) Energy coupling factor (ECF) Potassium transporter (KtrAB) Glutamate transporter (GltTk) Cytidine-Diphosphate diacylglycerol (Cds) Membrane-bound protease (PCAT) Folate transporter (FolT)
Copy #97
166 63 29
130 103 136 126 117 148 41 50 57
1341,397
Simulating Large Complex Membrane Systems
• 200nm diameter spherical vesicle constructed from 390k POPC lipid molecules
• Solvated with explicit water with 150mM NaCl salt concentration
• 1600 proteins, 400 copies each of:- Kv1.2 potassium channel [the small, inward-
pointing protrusion]- F1c10 ATPase complex [large, inward-
pointing, ball protrusion]- Multi-drug transporter P-glycoprotein (P-gp)
[the V-shaped protein]- Human glucose transporter (GLUT1)
Simulating Large Complex Membrane Systems
• 200nm diameter spherical vesicle constructed from 390k POPC lipid molecules
• Solvated with explicit water with 150mM NaCl salt concentration
• 1600 proteins, 400 copies each of:- Kv1.2 potassium channel [the small, inward-
pointing protrusion]- F1c10 ATPase complex [large, inward-
pointing, ball protrusion]- Multi-drug transporter P-glycoprotein (P-gp)
[the V-shaped protein]- Human glucose transporter (GLUT1)
Simulating Large Complex Membrane Systems
• 200nm diameter spherical vesicle constructed from 390k POPC lipid molecules
• Solvated with explicit water with 150mM NaCl salt concentration
• 1600 proteins, 400 copies each of:- Kv1.2 potassium channel [the small, inward-
pointing protrusion]- F1c10 ATPase complex [large, inward-
pointing, ball protrusion]- Multi-drug transporter P-glycoprotein (P-gp)
[the V-shaped protein]- Human glucose transporter (GLUT1)
Simulating Large Complex Membrane Systems
• 200nm diameter spherical vesicle constructed from 390k POPC lipid molecules
• Solvated with explicit water with 150mM NaCl salt concentration
• 1600 proteins, 400 copies each of:- Kv1.2 potassium channel [the small, inward-
pointing protrusion]- F1c10 ATPase complex [large, inward-
pointing, ball protrusion]- Multi-drug transporter P-glycoprotein (P-gp)
[the V-shaped protein]- Human glucose transporter (GLUT1)
Membrane Budding/Fusion
Workflow for Multi-Scale Modeling
Visual Molecular Dynamicstransform to potentials
Time Domain
construct mesh from shape(s)
123D Design
manipulate meshMeshLab
Workflow integrating multiple toolsdesigned to generate potentials for multi-scale simulations
Workflow for Multi-Scale Modeling
}remove sphere
points below planeremove plane points
within the sphere
planesphere
Grid Design and Construction
Membrane Budding/Fusion
Workflow for Multi-Scale Modeling
www.ks.uiuc.edu/~rcbernardi
There’s a dearth of new antibiotics to treat what the U.S. Centers for Disease
Control calls “nightmare bacteria.”
Bacterial Infection (MRSA)Methicillin Resistant Staphylococcus aureus
Staphylococcus bacterium
Bacterium Adhesin
Human Extracellular Matrix
Steered Molecular Dynamics
• Biased Simulation• Constant Force• Constant Velocity
- Pulling with a spring (Hook’s Law)‣ F = -k . Dx
- Atomic Force Microscopy (AFM)
Force
ExtensionH Grubmüller, et. al. Science (1996)S Izrailev, et. al. Langmuir (1997)SMD on NAMD Tutorial: http://www.ks.uiuc.edu/Training/Tutorials/namd/namd-tutorial-unix-html/node18.html
“Dummy Atom” Pulling Selection - Single Atom or Selection’s Center of Mass
Adhesion Mechanism – Staphylococcus epidermidis’ SdrGSdrG (serine-aspartate repeat protein G)
Targets Human’s Fibrinogen β (Fgβ)
BacteriumB1&B2domains
BacteriumSdrG
HumanFgβ
HumanExtracellularMatrix
Bloodflow
hosttargets
Staph.
Hermann Gaub (LMU)
Milles et.al, Science (2018)https://www.ks.uiuc.edu/~rcbernardi/
HermannGaub(LMU)
Adhesion Mechanism – Staphylococcus epidermidis’ SdrGSdrG (serine-aspartate repeat protein G)
Milles et.al, Science (2018)https://www.ks.uiuc.edu/~rcbernardi/
Over 2400 Steered Molecular Dynamics Simulations
Receptor:Li
Receptor:Li
https://www.ks.uiuc.edu/~rcbernardi/
Force Profile Sampling Loading Rate Dependent
Dudko,et.al.PhysicalReviewLetters(2006)Bullerjahn,et.al.NatureCommunications(2014)
Verdorferetal.JACS(2017)Milles et.al Science (2018)
• NAMD Enabled the Generation of Extensive Sampling (2400 Independent Simulations) • The Experimental Traces Agreed with the Simulation (Force Loading Rate Dependent)
- Dudko-Hummer-Szabo (DHS) Theory
• Enhanced Sampling Techniques• Replica Exchange Simulations
• Temperature• Solute Tempering• Bias Exchange• Grid Potentials
• String Method with Swarm of Trajectories
Replica Exchange
Metadynamics
Umbrella SamplingStochastic Simulations
• Biased Simulations• Steered MD• Target MD• ABF• Colvars - Wicked Useful
• RMSD• Distance• Orientation• Umbrella Sampling• Metadynamics . . .
Overcoming Timescale LimitationsWhen “Let It Go” Is Not an Option
gABF
constant-pH MD
TI
scripted variables
MtD
FEP
ABFTI / geometric
transformations Colvars
FEP / WCA
Hamiltonian hopping / FEP
eABF
US
ABF
MW/ABF
FREE-ENERGY METHODS
egABF
eABF
Roux group
Chipot group
Fiorin / Hénin
Others
2000
2001
2007
2016
2004
2011
2010
2016
2008
2009
20152014
2017
INTRODUCTION TO FREE-ENERGY CALCULATIONS INTRODUCTION
meta-eABF
2018Introduction to Free Energy Calculations (Chris Chipot) -
https://youtu.be/LCKtsR1ijsASLIDE COURTESY OF CHRIS CHIPOT
Use string method toidentify low-energytransition path andpartition space intoVoronoi polygons
Run many trajectories,stop at boundary
Faradjian and Elber. J. Chem. Phys. (2004)Bello-Rivas and Elber J. Chem. Phys (2015)Ma and Schulten JACS (2015)
Portable Innovation using Tcl and ColvarsMilestoning
Portable Innovation using Tcl and Colvars:Milestoning
Use string method toidentify low-energytransition path andpartition space intoVoronoi polygons
Run many trajectories,stop at boundary
Faradjian and Elber. J. Chem. Phys. (2004)Bello-Rivas and Elber J. Chem. Phys (2015)Ma and Schulten JACS (2015)
• Anton requires ~600 to 15000 ns MD for a single binding event (predicted kinetics are less accurate)
• Provides accurate & efficient binding kinetics predictions• Computes on- and off- rates
Votapka & Amaro, PLOS Comp Biol (2015)Votapka, Jagger, Heyneman, Amaro, J Phys Chem B (2017)SEEKER: https://github.com/nbcrrolls/SEEKR
Combines NAMD (MD) with BrownDye (BD) through milestoning to efficiently predict kinetics of ligand-receptor binding and off-rates
Portable Innovation using Tcl and Colvars:SEEKER
Using Colvars To Explore Transition Pathways in ABC Exporters
M. Moradi and ET PNAS (2013).M. Moradi and ET JCTC (2014).
M. Moradi and ET PNAS (2013) NBD Doorknob Mechanism
M. Moradi and ET PNAS (2013)M. Moradi and ET JCTC (2014) M. Moradi, G. Enkavi, and ET Nature Comm. (2015)Complex Reaction Pathways - https://youtu.be/ax74TgWY3wA and https://youtu.be/3hCaeFF05Jc
Complex Processes Require Complex Treatments
12 replicas x 40 ns (H1/H7) 50 replicas x 20 ns (10 Hs)
12 replicas x 40 ns (H1/H7) 24 replicas x 20 ns (H1/H7)
200 replicas (2D) x 5 ns 50 replicas x 20 ns
30 r x 20 ns 30 r x 20 ns 30 r x 20 ns
30 r x 20 ns 30 r x 20 ns
150 replicas
Describing a Complete Cycle (Adding Substrate)Requiring a Combination of Multiple Collective Variables
M. Moradi and ET PNAS (2013)M. Moradi and ET JCTC (2014) M. Moradi, G. Enkavi, and ET Nature Comm. (2015)Complex Reaction Pathways - https://youtu.be/ax74TgWY3wA and https://youtu.be/3hCaeFF05Jc
QM/MM – Fully Featured and Flexible Hybrid Interface
Main Features
CollaboratorsFrank Neese (ORCA)
Max Planck Institute forChemical Energy Conversion
Mülheim an der Ruhr, Germany
Gerd Rocha (MOPAC-GPU)Theoretical Quantum Chemistry Group
UFPB - João Pessoa, Brazil
• Interface to ORCA, MOPAC, and “Generic Interface to any QM Software”• Multiple QM/MM Coupling Schemes (Charge Redistribution Schemes)• PME for Long-Range Electrostatics• Solvent Molecule Switcher (necessary for long timescale QM/MM simulations)• Multiple Independent QM Regions• Easy Setup Interface with QwikMD• Mix and Match (Combine with any other NAMD Feature):
-Polarizable Force Field in MM Region-Replica Exchange Molecular Dynamics-Adaptive Biasing Force-Steered Molecular Dynamics-and many other features.
• Multi-Level QM Regions (Coming Soon)
QM/MM highly requested by usersIntegrating Quantum Mechanics (QM) to
Molecular Mechanics (MM), QM/MM allows study of chemical reactions and many other
quantum processes.
https://www.ks.uiuc.edu/Research/qmmm/
Enhanced Sampling Techniques can be used with QM/MM to investigate e.g. reaction mechanisms
Combining the Collective Variables module of NAMD with QM/MM allows for the investigation of reaction pathways with the utmost level of details.
Melo et al. Nature Methods (2018)Bernardi et al., Biochimica et Biophysica Acta (BBA), (2015)
Stochastic Simulations Replica Exchange Molecular Dynamics
Metadynamics
Combining Hybrid QM/MM Simulations with Enhanced Sampling
NAMD Constant pH Simulations
http://www.ks.uiuc.edu/Training/Tutorials
• Protein Residues• neMD/MC
Molecular Dynamics
• Dual topology • TCL scripts
- Milestoning - and others...
Staph nuclease (SNase)
NAMD Constant pH simulations
• Drive alchemical growth with nonequilibrium work
• Accept/reject with a generalized Metropolis criterion
MC sample of auxiliary coordinates
removal of auxiliary coordinates
neMD alchemical growth
λ=1/3λ=0
λ=2/3
λ =1
Stern J Chem Phys, (2007)Chen and Roux J Chem Theory Comput (2015)
Radak, et al. J Chem Theory Comput ( 2017)
Electrostatics with Multilevel Summation Method
• Uses Hierarchical Interpolation of Smoothed Pairwise Potential
• Improves Parallel Communication over FFT-based PME
Simulate non-periodic systems Simulate semi-periodic systems
Hardy, et al. JCTC (2015)Hardy, et al. JCP (2016)
NAMD Coming Attractions
• Fast Single Node GPU-Accelerated Mode• Improved single-GPU/single-node performance• GPU support for FEP & TI• GPU support for multilevel summation method
(MSM)• GPU support for Drude polarizable force field• Long-range dispersion forces (LJ-PME, LJ-MSM)• Support for Martini 2.x force field
NAMD – Single-node GPU Performance TraceCPU integrator is now the bottleneck
1% of computation is now ~50% of timestep work.
Next step: Offload it to the GPU
NVIDIANsightSystemsprofiler
Integrator running on CPU cores
Force compute running on GPU
Ongoing Development
GPU integrator per patch
GPU integrator per
CPU core
GPU integrator per system (Ongoing)
Benchmarked System: ApoA1 (92.224 atoms)2fs timestep12A cutoff
Intel Xeon E5-2650 (16 Cores) + Nvidia Titan V
1.2 ns/dayOne kernel launch per patch
Lots of memory transfers
14.5 ns/dayBetter usage of GPU resources
Still bottlenecked by memory transfers
45.5 ns/dayAvoiding memory transfers
Keeping data on GPU as much as possible
NanosecondsperDay
0
12.5
25
37.5
50
PerPatch PerCore PerSystem
45.5
14.5
1.2
NAMD – Single-node GPU Performance Trace
Future Developments
Gaps in the GPU usage — get rid of additional Charm++ related overheads
Atom migration steps are becoming larger as intermediate steps shrink — move them to the GPU as well
NAMD – Single-node GPU Performance Trace
Thanks to: NIH, NSF, DOE, NCSA, ALCF, OLCF,and 20+ years of NAMD and Charm++ developers and users.
João V. RibeiroNIH Center for Macromolecular Modeling and BioinformaticsUniversity of Illinois at Urbana-Champaign http://www.ks.uiuc.edu/Research/namd/