HPC-Sig 2018: Sampling Molecular Conformations and Dynamics in Virtual Reality

Overview of the Nano Simbox VR platform for interactive molecular dynamics.

Transcript

1. Sampling Molecular Conformations and Dynamics in Virtual Reality Mike O’Connor Dept. of Computer Science & Chemistry mikeoconn.org
2. This talk • A quick history of molecular modelling. • Combining virtual reality and GPU-accelerated simulations for real- time molecular modelling. • Outstanding challenges. • Example applications. • Demonstration!
4. The VR team
5. Becky Sage, Phill Tew, Mark Wonnacot, Becky Davies, Jenny Blackwood, Liz Rabone, Balazs Hornung, Liam Cullingford, Alix Briskham
6. We span a lot of disciplines… Super-Computing Human Computer Interaction Machine Learning Molecular Dynamics Digital Arts
7. Molecular Models: A Brief History
9. Molecular models in 1958 were indeed frozen out and had little choice but to flee, and it took three decades for every journal to insist on depositing atomic coordinates as a prerequisite for publication. From structure to protein folding Solving the structures of proteins at high resolution uncovered a new problem and initiated a novel field of research — that of protein folding. It was known from seminal experiments by Christian Anfinsen that small proteins could spontaneously refold from their denatured states, and so the primary structure (sequence) of a protein dictates its tertiary structure9. The ‘protein folding prob- lem’ consists of two parts: first, the tertiary structures of proteins need to be predicted from their primary sequences, and second, the pathway of folding and unfolding must be predicted. Cyrus Levinthal and others, such as Michael Levitt and Oleg Ptitsyn, wanted to predict tertiary structures by predicting their folding pathways. Levinthal famously pointed out in almost offhand remarks during a meeting in 1969 (REF. 10) that it seemed impossible that an unfolded protein could fold spontaneously by a random process on a biological time scale. Mechanisms were proposed that could overcome the ‘Levinthal paradox’ by simplifying the folding process and breaking it down into subprocesses that could occur stepwise. Ptitsyn suggested Figure 1 | Three-dimensional structure of myoglobin. The low-resolution structure of myoglobin that was published by John Kendrew and colleagues in 1958 (REF. 1). This figure in the Nature paper was reconstructed by the author using the original figures in the archives of the Medical Research Council Laboratory of Molecular Biology, Cambridge, UK. Polypeptide chains are in white and the grey disc represents the haem group. The three spheres show positions at which heavy atoms were attached to the molecule (black, Hg of p-chloro-mercuri-benzene-sulphonate; dark grey, Hg of PERSPECTIVES Alan Fersht, “From the first protein structures to our current knowledge of protein folding: delights and scepticisms,” Nature Rev Mol Cell Biology, 2008
10. “Hello molecular model!”
11. Michael Levitt’s 1968 model of hen egg white lysozyme
12. Structure is Only Part of the Story – Proteins Are Dynamic Objects “Everything that living things do can be understood in terms of the jigglings and wigglings of atoms” -Richard Feynman
13. 1972: The first movie of a molecular dynamics trajectory Dynamics of Pancreatic Trypsin Inhibitor
14. Molecular Simulation: The next 40 years… Computers have gotten more & more powerful… Simulations have gotten better and better Graphics consoles have gotten better & better “Human-machine interaction is usually limited to a seated man poking at a machine with his fingers or perhaps waving his hands over a data tablet…” Myron W. Krueger, “Responsive Environments”, National Computer Conf., 1977
15. The limitations of physical models “Trust me. This protein normally has secondary structure!”
16. Nano Simbox VR
17. O’Connor, Deeks et al., arXiv:1801.02884, Science Advances, in press
18. O’Connor, Deeks et al., arXiv:1801.02884, Science Advances, in press
19. Modular Design • Generic server/client model. Does not have to be a molecular simulation! • Connect to existing simulation codes: • GROMACS • OpenMM • LAAMPS • DFTB+ • DL-POLY
20. Current Setup: Local Dev & Studies • VR => Small scale HPC in the lab. • Run small/medium simulations directly on our VR machines. • Larger simulations run over LAN on our workstation with 4 GPUs. • We currently avoid multi-node. • Would like to try running on Bristol / national HPC. • Can of worms I’ve been avoiding opening! • Complex software stack. Containers? • Follow-up simulations run on HPC (BCP3/4).
21. Current Setup: Education, Public Access • Cloud development being led by Interactive Scientific. • Avoids having to worry about client IT infrastructure. • Scalable deployment on Oracle cloud (EU, US East, US West). • Containers on bare metal servers. • Currently no GPUs => small molecular systems.
22. Uses for the Nano Simbox VR • Simulation building & preparation. • Analysis of completed simulations. • Remote visualization of live simulations. • Interactive steering of simulations.
23. Interactive MD (IMD) • Previous efforts: • Stone et al., 2001 • Dreher et al., 2013 • Use haptic devices. • Problems: • Colocation • Latency • Availability • Size of simulation • Simulation timescales 23
24. VR: It’s just a gimmick, right?
25. O’Connor, Deeks et al., arXiv:1801.02884, Science Advances, in press
26. https://isci.itch.io/nsb-imd Download our software and try it yourself!
27. Remote Visualization in VR: Challenges at the Intersection between HCI and HPC
28. Simulation Size is Not (Always) Our Bottleneck 0.125 0.25 0.5 1 2 4 8 16 32 64 4,096 16,384 65,536 262,144 1,048,576 Integration Step / ms Number of Atoms 30 FPS 60 FPS OpenMM OpenMM IMD Current VR Rendering Limit!
29. Cloud Latency 0 0.1 0.2 0.3 0.4 0.5 0.6 10 30 50 70 90 110 130 150 170 190 210 Probability time /milliseconds bristol-frankfurt cable bristol-ashburn cable bristol-phoenix cable
30. Throughput Limits
31. Optimisation: Molecular Level Of Detail
32. Some More Applications
33. O’Connor, Deeks et al., arXiv:1801.02884
34. Scientific ‘gamification’ Casting scientific problems as games that scientific players can solve Which computer is faster?
35. A quantum chemistry VR game
36. Can we design algorithms that let machines learn from the experts?
37. Conclusions • New technologies lets us revisit older and more immersive domains of molecular modeling, allowing us to interact with living & breathing structures animated by rigorous physics • There’s lots of interesting opportunities to explore using this technology, in both research & also in science education • User-controlled HCI studies testing various aspects of this framework are encouraging • There are lots of challenges remaining in optimising for remote visualization.
38. Thanks! mikeoconn.org m.oconnor@bristol.ac.uk @mikeoconnor0308
39. Thanks very much to the funders
40. Molecular Simulation: Sampling Problems • Transitions between minima are rare. • Brute force simulation is not fast enough. • Simulations run for µs*, events happen on ms or seconds. • The rare events are often what is interesting! Source: Thomas Splettstoesser CC BY-SA 3.0