With over nine million infected and nearly half a million dead, the COVID-19 pandemic has seized the world’s attention for several months. It has also dominated the supercomputing sector, with COVID-related research receiving major allocations on nearly every research supercomputer in the world (and many industrial supercomputers). It’s not surprising, then, that at ISC 2020, the virtual conference opened, revealed the new Top500 list – and then got straight to COVID-19.
In the focus session, three speakers addressed how HPC is fighting back against the coronavirus: Satoshi Matsuoka of RIKEN, which just nabbed the top spot in the Top500 with its Fugaku system; Peter Coveney of the Centre of Excellence in Computational Biomedicine, which is working to make HPC and machine learning actionable in a rapid drug development workflow; and Rick Stevens of Argonne National Laboratory, where researchers are working with the COVID-19 HPC Consortium to comb through billions of molecules.
Satoshi Matsuoka highlights Fugaku’s role in COVID-19 research
Fugaku, the most powerful supercomputer in the world, is in business early thanks to COVID-19. The system – situated at Riken in Japan – was scheduled to launch in 2021. When the pandemic struck, plans changed. “It was decided by [Japan’s] Ministry of Education, Culture, Sports, Science and Technology that we will utilize [not only supercomputers that are already available] but also [deploy Fugaku early], almost a year ahead of schedule, to combat COVID-19,” explained Matsuoka, director of the Riken Center for Computational Science (R-CCS).
Fugaku’s showstopping 415 Linpack petaflops are close to triple the performance of the runner-up, Oak Ridge’s newly dethroned Summit system. At 158,976 nodes, Fugaku is the largest system ever created in terms of nodes, footprint and power consumption. The software, Matsuoka said, is “quite standard,” allowing for broad usability without much Fugaku-specific tweaking.
“They’re largely divided into two areas,” Matsuoka said of Fugaku’s COVID-19 workloads. “One is medical-pharma – so trying to see how the virus behaves, what are the effective drugs, especially how we can repurpose existing drugs and so forth and also how a vaccine is made. So these are molecular-level investigations of the virus and its countermeasures. The other is more macroscopic – so we’re trying to see how these viruses are transmitted and what are the mitigation measures and how it will impact society.”
Matsuoka highlighted several of the COVID-19 projects taking advantage of Fugaku’s early arrival. One Riken researcher, for instance, is studying conformational changes of the spike protein using a highly scalable molecular dynamics code. Another researcher is using fragment molecular orbital calculations to investigate the energy levels of the spike protein, scaling across hundreds of thousands of Fugaku’s CPUs. “On [Fugaku’s predecessor] the K computer,” Matsuoka said, “this calculation would have taken days, weeks, multiple weeks – on Fugaku, … they have been able to do this in just three hours.”
Other researchers are using Fugaku to run socially oriented simulations, such as simulating droplets in indoor spaces like trains or simulating the spread effects of using face masks or contact tracing applications, Matsuoka said – and, of course, there are more to come. “So if you have any good ideas,” he said, “go to the website and you can apply.”
Peter Coveney describes a new, HPC- and AI-driven model for drug development
Coveney, the second speaker, runs the Centre of Excellence in Computational Biomedicine (CompBioMed), an initiative funded by the European Union that is currently redirecting its research efforts and computational research to the study of and drug development for COVID-19. Coveney (who also teaches at University College London) stressed the need to “invert the model [of drug development] as it currently exists” using advanced IT.
“The opportunities there are enormous,” Coveney said. “What we’re really trying to do is transform the approach to biomedicine, to be able to move it from a highly empirical approach … to putting a priority on the predictions that come out of computers.”
But to do that, he said, the computational results had to be actionably accurate – and perhaps even more difficult, they had to be quickly produced. Molecular screening, however – the crux of computational drug design, whereby compounds are fitted to targets on the virus’ proteins – is labor-intensive, time-consuming and expensive ($1 to $10 a compound, with billions of compounds to screen for COVID-19).
Coveney outlined how CompBioMed worked with over 40 partners around the world to streamline the computational drug design pipeline. CompBioMed gained access to a wide range of supercomputers, from SuperMUC-NG (the most powerful supercomputer in the EU) to Piz Daint, Archer, Summit, Frontera, Theta and more. The researchers used a piece of middleware called Radical Cybertools to run workflows across a large number of nodes on multiple machines.
With computing power in hand, CompBioMed focused on how to ensure “validation, verification and uncertainty quantification” (or “VVUQ”) in the pipeline. “This is designed in general to raise confidence in HPC simulation,” Coveney said.
To effectively leverage the computing power and ensure “VVUQ,” CompBioMed combined machine learning with molecular dynamics. Machine learning was used first to whittle down the near-infinite list of candidate molecules. “We have to do searches in a hurry,” Coveney said. “We want to use computationally very fast methods that are also cheap … to search huge libraries of molecules, to explore chemical space, to predict new molecules and so on.”
Then, with the list whittled down, CompBioMed used molecular dynamics simulations – 20 to 30 of them at a time. As Coveney explained, a single molecular dynamics simulation could have a large number of errors. “But if you run many of them concurrently … we can run those on very large supercomputers all at the same time,” Coveney said. “Then we can make reliable predictions that get fed back to another stage of the machine learning.”
The best candidate compounds from this process are then submitted to medical research labs for further testing. “We are already discovering many tens to hundreds of potential compounds that can be investigated by our experimental colleagues,” Coveney said. “And indeed, that’s happening already.”
“We’re trying to change the way medicine is actually understood and applied,” Coveney concluded. “We want to make the subject more amenable to scientific investigation, that it should revolve around theory, modeling and simulation in addition to experimental research.”
Rick Stevens dives into the COVID-19 HPC Consortium and machine learning-enabled research
Finally, Stevens took the virtual stage. Stevens – associate laboratory director at Argonne National Laboratory – has been working closely with the COVID-19 HPC Consortium, a public-private effort to pool supercomputing resources for COVID-19 research. Currently, the effort has over 40 members, comprising some 483 petaflops of resources, 50,000 GPUs, 136,000 nodes and five million CPU cores.
As Stevens explained, the projects being tackled by the consortium fall into three broad categories: first, basic science, including things like analyzing the virus’ structure, protein functions and virus evolution; second, therapeutics (“the largest group”), aiming to discovery drug targets on the virus, design drugs and discover repurposable drugs; and finally, patient care – “things more related to optimizing the healthcare system or epidemiology.”
Stevens outlined some of the key work, especially where it intersected with Argonne. “If you’re gonna work on this problem, you need to understand the enemy,” Stevens said, describing how Argonne has used its Advanced Photon Source (APS) to identify new structures of COVID-19, which in turn produce new drug targets for simulations to examine.
Like Coveney, Stevens highlighted the intersections of AI and supercomputing as viable pathways for processing massive amounts of compounds in a relatively short time frame. For instance, he said, researchers were using AI to reconcile models of proteins from various sources to produce even more accurate models. In the spring, Argonne also began assembling a large database – around 60 TB – containing descriptors, images and more for over four billion compounds, with the aim of producing massive datasets for machine learning applications.
“One of the strategies that we have is to use a combination of high-throughput virtual docking … to generate scores – generate them on thousands or millions of data points,” Stevens said, “but then use that data to train machine learning models and do inference on a much larger scale.” As in Coveney’s research, the most promising hits are then sent for wet lab screening.
Stevens also discussed the use of machine learning to understand the “trajectories” of molecular dynamics simulations and the use of reinforcement learning to essentially build drug molecules from the ground up, adding to them iteratively to improve the docking score.
“One of the overall challenges here, of course, is that there’s over 10⁶⁰ possible drugs,” Steven said, “and you can only test at the end of the day, in humans, a small fraction of these.” But now, with AI and supercomputing converging to create a new model of rapid drug design, that might be enough.