2021 Victorian GPGPU Research Symposium

Computational insights into the origin of SARS-CoV-2

The devastating impact of the COVID-19 pandemic caused by SARS–coronavirus 2 (SARS-CoV-2) has raised important questions about which species may be susceptible to the virus and politically charged questions about its origin. Companion or commercial animals may act as SARS-CoV-2 reservoirs for the virus and seed outbreaks as was seen recently with minks in Scandinavia. To throw some light onto these issues, we undertook an in silico structural homology modelling, protein–protein docking, and molecular dynamics simulation study of SARS-CoV-2 spike protein’s ability to bind angiotensin converting enzyme 2 (ACE2) from relevant species. The study was conducted using extensive CPU and GPU resources provided by Oracle Cloud Systems.

Surprisingly, the spike protein exhibited the highest binding to human (h)ACE2 of all the species tested, Interestingly, pangolin ACE2 showed the next highest binding affinity despite having a relatively low sequence homology, whereas the affinity of monkey ACE2 was much lower despite its high sequence similarity to hACE2. ACE2 species in the upper half of the predicted affinity range (monkey, hamster, dog, ferret, cat) have been shown to be permissive to SARS-CoV-2 infection, These findings show that the earliest known SARS-CoV-2 isolates were surprisingly well adapted to bind strongly to human ACE2, helping explain its efficient human to human respiratory transmission. This study highlights how computational methods can rapidly generate information on novel viruses to aid in countermeasure development.

---

Dave has an unusually broad formal training in chemistry, physics, chemical engineering and radioastronomy. He is a Professor of Biochemistry & Genetics at La Trobe Institute for Molecular Science at La Trobe University, an adjunct Professor of Medicinal Chemistry at the Monash Institute for Pharmaceutical Sciences, a visiting Professor in Pharmacy at the University of Nottingham, and a Fellow in evolutionary robotics at CSIRO Data61. He previously spent over 30 years at CSIRO researching the application of computational chemistry, AI, and machine learning methods to the design of drugs, agrochemicals, nanomaterials and biomaterials. He is ranked 227th out of 81,000 medicinal chemists, and 999th out of 520,000 chemists worldwide (Mendeley 2019). He has authored over 250 refereed journal articles and book chapters, has an H index of 51, and is an inventor on 25 patents. He has won several prestigious awards including the CSIRO Medal for Business Excellence, RACI’s Adrien Albert award for contributions to medicinal chemistry, the ACS Herman Skolnik award for excellence in cheminformatics, and a Royal Academy of Engineering (UK) Distinguished Fellowship (bioengineering). He is past President of the Federation of Asian Chemical Societies (FACS) and of the Asian Federation for Medicinal Chemistry (AFMC), and past Chairman of the Royal Australian Chemical Institute.

GPU software infrastructure for ML molecular simulations

In this talk, I will discuss software development efforts in our lab to support GPU-accelerated machine learning tasks for computational chemistry and molecular simulations.

---

The Isayev lab works at the interface of theoretical chemistry, pharmaceutical sciences and computer science. In particular, we are using molecular simulations and artificial intelligence (AI) to solve hard problems in chemistry. We are working towards the acceleration of molecular discovery by the combination of AI, informatics and high-throughput quantum chemistry. We also focus on both generative and predictive ML models for chemical and biological data.

The GPGPU developer experience has a long way to go

Many people now use GPGPU every day, mainly through Python with the PyTorch and TensorFlow libraries. Other more recent approaches include Jax and Julia. A relatively small number of programmers use CUDA or similar low-level approaches directly from C or FORTRAN. However, all of these options have significant deficiencies when it comes to the developer experience. For PyTorch, TensorFlow, and Jax, these deficiencies may be unfixable. It's time for a conversation about what developers will need from GPGPU in the coming years.

---

Jeremy Howard is a data scientist, researcher, developer, educator, and entrepreneur. Jeremy is a founding researcher at fast.ai, a research institute dedicated to making deep learning more accessible. He is also a Distinguished Research Scientist at the University of San Francisco, the chair of WAMRI, and is Chief Scientist at platform.ai.

Previously, Jeremy was the founding CEO Enlitic, which was the first company to apply deep learning to medicine, and was selected as one of the world’s top 50 smartest companies by MIT Tech Review two years running. He was the President and Chief Scientist of the data science platform Kaggle, where he was the top ranked participant in international machine learning competitions 2 years running. He was the founding CEO of two successful Australian startups (FastMail, and Optimal Decisions Group–purchased by Lexis-Nexis). Before that, he spent 8 years in management consulting, at McKinsey & Co, and AT Kearney. Jeremy has invested in, mentored, and advised many startups, and contributed to many open source projects.

He has many media appearances, including writing for the Guardian, USA Today, and the Washington Post, appearing on ABC (Good Morning America), MSNBC (Joy Reid), CNN, Fox News, BBC, and was a regular guest on Australia’s highest-rated breakfast news program. His talk on TED.com, “The wonderful and terrifying implications of computers that can learn”, has over 2.5 million views. He is a co-founder of the global Masks4All movement.

GPU-powered Natural Language Processing: Opportunities and Challenges in Academic Research

GPUs are practically prerequisite for current research in natural language processing (NLP), partly due to the infamous success of pre-trained transformer language models such as GPT-3. In this talk I will briefly explain their architecture and showcase their utility by the example of our recent work on human-like language understanding. I will conclude with a brief discussion of some challenges of GPU-powered NLP research in academia, with an eye towards equity and sustainability.

---

Lea Frermann is a Lecturer in Natural Language Processing in the School of Computing and Information Systems at The University of Melbourne.

Her research focuses on efficient and robust learning and inference in the face of the complexity of the world, as approximated for example through large corpora of child-directed speech, or the plots of books and films. She has worked automatic acquisition and representation of common sense knowledge; on inducing structured representations of storylines for improved summarisation and question answering; and is interested in how stereotypes and biases in our society are reflected in (and can be removed from) models of natural language.

She received her M.Sc. from Saarland University in Germany in 2013, and her Ph.D. from the University of Edinburgh in 2017. She was a research associate at the University of Edinburgh, and an applied scientist at Amazon AI (Berlin) prior to joining Melbourne University in July 2019.

GPU-enabled high-fidelity CFD with industrial impact

CFD predictions are becoming increasingly important in the design of aircraft engines because correlation-based methods are unable to further improve efficiency and laboratory experiments with the required fidelity are prohibitively expensive. This presentation will show how physical insight relevant to designers can be extracted from high-fidelity simulations enabled by GPU acceleration. It will also discuss recent high-fidelity simulations of performance-enhancing surface textures, and describe how our research can both respond to manufacturer concerns about their practical use and potentially inform future design.

---

Dr Melissa Kozul completed her PhD at the University of Melbourne in 2018. From 2018 until early 2021 she was a Postdoctoral Fellow within the Thermo Fluids Research Group at NTNU in Norway, before returning to the University of Melbourne as a Research Fellow in the Department of Mechanical Engineering.

Melissa’s research expertise is in the high-fidelity simulations of fundamental turbulent flows that feature critically in energy and transport technologies. For this she often employs strategically-designed numerical ‘thought experiments’ to decouple effects that are conflated, or add controlled physics. To date her research interests have included wall-bounded turbulent flows, homogeneous isotropic turbulence and turbulent multiphase flows.

Optimising DL pipelines to take advantage of NVIDIA DGX stations: Case Study: Fungal infection detection in CT scans.

Fungal infections are a global disease that affects more than a billion people worldwide. Invasive infections with mold, including fungal pneumonia, affect about 250,000 to 300,000 people worldwide and is life-threatening with a 30-100% mortality rate. This project looks at using AI on CT scan data to detect fungal infections in the lungs. We use this complex problem as a case study to see how we applied state-of-the-art Deep learning frameworks to accelerate our experimentation on HPC hardware. In this presentation, we lay out the challenges involved and discuss our approach to the problem on how to better optimise experiments to take full advantage of NVIDIA DGX systems.

---

Suhag Byaravalli Arun is a research DevOps specialist with an interest in MLOps. He previously worked on various aspects of Data Science/ML problems from fast-prototyping, to productionizing and maintaining solutions, primarily specializing in chipping away the "Hidden Technical Debt in Machine Learning".

GPUs and AI for Image the Eye

Cylite is a technology development company, creating the next generation of imaging and metrology systems for ophthalmic and related markets. Analyzing the volumetric images that their instruments produce in a clinically relevant time-frame is a critical and challenging task. This presentation will provide an overview of how both GPUs and machine learning are crucial components of meeting this goal.

---

Dr Stephen Moore is a senior software engineer at Cylite. His major responsibilities include the design and implementation of the back-end analysis software that accompanies their 'Hyperparallel' OCT imaging systems, designed for use by optometrists and ophthalmologists. The analysis process itself encompasses a vast amount of processing, which can only be performed in a clinically relevant timeframe by leveraging the compute power of modern GPUs. Such tasks include the raw signal processing required to produce OCT volumetric data describing the eye, application of various convolutional neural networks for certain filtering and classification tasks, and image registration/resampling techniques, that are all needed for the derivation of clinical metrics.

Ensemble Kalman filter estimation of ocean optical parameters in numerical simulations of the global climate

Coupled general circulation models (GCM) of the atmosphere, ocean, land and sea-ice have many uncertain parameters, which contributes to inherent model biases. To address this problem we use the CSIRO Climate re-Analysis and Forecast Ensemble (CAFE) system to estimate both the climate state (atmosphere, ocean, sea-ice) and also time varying parameter maps controlling the ocean heat penetration depth. The CAFE system adopts a 96 member ensemble transform Kalman filter within a strongly coupled data assimilation (DA) framework, which estimates the parameters (and states) by minimising the error between short term DA cycle forecasts of the climate model and a network of real world atmospheric, oceanic, and sea-ice observations. The DA estimation code (enkf-c) and climate model (GFDL CM2.1) and are both parallelised via spatial domain decomposition using MPI, with the propagation of the climate models embarrassingly parallel along the ensemble member dimension. This approached produced an improved fit to observations over the period from 2010 to 2012, and also a systematically reduced bias in multi-year climate forecasts during the out-of-sample period from 2012 to 2020.

---

Dr Vassili Kitsios completed a PhD with the University of Melbourne and the Université de Poitiers on fluid dynamical stability and model reductionof aerospace flows (2006-2010). He then undertook post-doctoral research with the CSIRO Oceans and Atmosphere division (2010-2013) and the Monash University Laboratory for Turbulence Research in Aerospace and Combustion (2013-2016),on the massively parallel numerical simulation(>32,000 cores)and stochastic parameterisation of atmospheric, oceanic and boundary layer turbulence. He then held an industrial mathematical finance position at a hedge fund (2016-2017) using time series analysis to develop trading algorithms on the basis of macroeconomic themes and market conditions. Since re-joining CSIRO in 2017, he has been undertaking research on the ensembledata assimilation and stochastic modelling methods for improved climate state / parameter estimation and forecasting. His most recent research involves quantifying the influence climate variability and change has on financial markets and health indicators.

Simulations of confined bluff body flows on GPU architectures

Bluff body flows are often considered in isolation with the assumption that nearby objects have little influence. However, in many engineering applications, such as buildings within cities, nearby objects will often have a bounding effect to the bluff body, and alter the nature of the flow.

In this talk, we will show how GPU acceleration has enabled for high fidelity simulations to be conducted in such a geometry and discuss the effects confinement has on transitional and turbulent flows past bluff bodies.

---

Wilson Lu is a PhD candidate in the Department of Mechanical Engineering at the University of Melbourne. His interests are in the dynamics of transitional and turbulent confined bluff body flows.

The adoption of GPGPUs by the scientific community has accelerated the progress of research in many disciplines. Computational chemistry and materials science have certainly embraced the technology and there are now many software packages with GPGPU capability. In this talk I will give a brief overview of the issues involved with GPGPUs - what to expect and what to watch out for - in the context of NCI and the Gadi supercomputer.

An Active Site Inhibitor Induces Conformational Penalties for ACE2 Recognition by the Spike Protein of SARS-CoV-2

The novel RNA virus, severe acute respiratory syndrome coronavirus II (SARS-CoV-2), is currently the leading cause of mortality in 2020, having led to over 1.6 million deaths and infecting over 75 million people worldwide by December 2020. While vaccination has started and several clinical trials for a number of vaccines are currently underway, there is a pressing need for a cure for those already infected with the virus. Of particular interest in the design of anti-SARS-CoV-2 therapeutics is the human protein angiotensin converting enzyme II (ACE2) to which this virus adheres before entry into the host cell. The SARS-CoV-2 virion binds to cell-surface bound ACE2 via interactions of the spike protein (s-protein) on the viral surface with ACE2. In this paper, we use all-atom molecular dynamics simulations and binding enthalpy calculations to determine the effect that a bound ACE2 active site inhibitor (MLN-4760) would have on the binding affinity of SARS-CoV-2 s-protein with ACE2. Our analysis indicates that the binding enthalpy could be reduced for s-protein adherence to the active site inhibitor-bound ACE2 protein by as much as 1.48-fold as an upper limit. This weakening of binding strength was observed to be due to the destabilization of the interactions between ACE2 residues Glu-35, Glu-37, Tyr-83, Lys-353, and Arg-393 and the SARS-CoV-2 s-protein receptor binding domain (RBD). The conformational changes were shown to lead to weakening of ACE2 interactions with SARS-CoV-2 s-protein, therefore reducing s-protein binding strength. Further, we observed increased conformational lability of the N-terminal helix and a conformational shift of a significant portion of the ACE2 motifs involved in s-protein binding, which may affect the kinetics of the s-protein binding when the small molecule inhibitor is bound to the ACE2 active site. These observations suggest potential new ways for interfering with the SARS-CoV-2 adhesion by modulating ACE2 conformation through distal active site inhibitor binding.

Here, we demonstrate the usefulness of the University of Melbourne high performance computing cluster (SPARTAN) in theoretically evaluating this mechanism for a potential cure for SARS-CoV-2 infection. The GPGPU cluster was shown to accelerate MD simulations of the s-protein RBD/ACE2 complex by approximately 3-fold, enabling rapid sampling of this construct to assist medicinal chemistry research.

---

Dr Billy Noonan is a post-doctoral researcher at RMIT University in Melbourne, working within the Materials Modelling and Simulation (MMS) research group. His current research within this group involves the use of molecular dynamics simulations to assist rational drug design, and to understand the self-assembly mechanisms of specific β-peptides. He completed his PhD at the Monash Institute of Pharmacy and Pharmaceutical Science (Melbourne) in 2020. Here, he specialised in the application of molecular dynamics, free energy calculations and enhanced sampling algorithms to assist medicinal chemistry projects. Completing an Honours in a carbohydrate chemistry and drug design at Griffith University’s Institute for Glycomics (Gold Coast) in 2012, Dr Billy Noonan also has some laboratory experience to complement his computational expertise.

Co-authors: Nevena Todorova, Ketav Kulkarni, Marie-Isabel Aguilar, and Irene Yarovsky

Prof. Walsh's research interests and expertise focus on molecular modelling, chiefly of interfaces, using molecular dynamics simulations and first-principles calculations.

After graduating with a B.Sci(Hons) from the University of Melbourne, Prof. Walsh earned her PhD degree in theoretical chemistry from the University of Cambridge, U.K., working in the group of Prof. David Wales (FRS) in the Dept. of Chemistry, as a Cambridge Commonwealth Trust scholar. She then joined the Dept. of Materials, University of Oxford, U.K. as a postdoctoral researcher in the Materials Modelling Laboratory (MML) with Prof. Adrian Sutton (FRS). Staying in the MML, she held a Glasstone Fellowship, in addition to a Junior Research Fellowship at Linacre College, University of Oxford. Following this, she joined the faculty of the University of Warwick, U.K., as a joint appointment in the Dept. of Chemistry and the Centre for Scientific Computing. In 2012 she returned to Australia on a veski Innovation Fellowship and joined the Institute for Frontier Materials at Deakin University in Australia, where she currently holds the position of Professor of Bio/Nanotechnology.

Prof. Walsh served on the Australian Research Council (ARC) College of Experts from 2015-2017, and currently serves on the ARC Medical Research Advisory Group.

A HPC Virtual Laboratory: AI/ML for Weather and Climate Science

In this presentation I will show that data driven models can predict important meteorological quantities of interest to society, such as global high resolution precipitation fields (0.25 degrees), sea surface temperatures and 500 hPa geopotential heights, and can deliver accurate forecasts of the future state of the atmosphere without prior knowledge of the laws of physics and chemistry. I will also show how these data driven methods can be scaled to run on super-computers with up to 1024 modern graphics processing units (GPU) and beyond resulting in rapid training of data driven models, thus supporting a cycle of rapid research and innovation. Taken together, these two results illustrate the significant potential of data driven methods to advance atmospheric science and operational weather forecasting.