Are supercomputers set to transform pharma R&D?
Tech company NVIDIA this year launched the UK’s most powerful supercomputer to help researchers solve pressing medical challenges. Darcy Jimenez learns more about the system.
Advanced technology is now commonplace in pharma, where innovations like artificial intelligence (AI) and machine learning accelerate and improve the accuracy of drug discovery and development efforts.
Add to the list the supercomputer – a powerful class of computer, vastly superior to general-purpose computers in terms of speed and performance, that are commonly used for data-intensive scientific purposes.
But what does this ultra-powerful technology mean for medicine?
US tech company NVIDIA this year launched Cambridge-1, the UK’s most powerful supercomputer, to enable UK healthcare researchers to solve pressing, large-scale medical challenges using powerful AI and simulation.
The product of a $100m investment by NVIDIA, Cambridge-1 will enable scientists to better understand brain diseases like dementia, use AI to design new drugs, more accurately identify disease-causing variations in human genomes, and more.
The projects will be pursued in collaboration with the computer’s first users: AstraZeneca, GlaxoSmithKline, Guy’s and St Thomas’ NHS Foundation Trust, King’s College London, and Oxford Nanopore Technologies.
Boosting the UK’s life sciences
Craig Rhodes, NVIDIA’S EMEA industry lead for AI for healthcare and life sciences, says medical research was the natural target for Cambridge-1’s processing capabilities.
“We’ve got enough data – sometimes too much data – but now computation is the bottleneck,” Rhodes explains. “So, what we wanted to do as part of our investment into the cure for Covid, and now looking at cures for cancer, was to remove that computational bottleneck.
“With all the data that we’ve got, we can do things much more accurately; we can produce results that are much more efficient and effective for dealing with clinical change, drug discovery change, and different clinical practices.
“Us deploying Cambridge-1 is showing clinical and drug discovery organisations and organisations doing large, national population sequencing that supercomputing – removing the computational bottleneck – will give them an answer much faster and much more accurately,” he says.
“If you take all of those dimensions: the UK being such a hotbed for AI, NVIDIA’s investment in the key organisations in the UK, as well as wanting to accelerate [finding cures for] Covid-19 and now cancer – all of those things put together is such a unique opportunity to NVIDIA to go and do something really, really good and worthwhile.”
“To look at microscopy data, look at the structure of compounds, identify how that compound might change subtly when it when it locks into a target, is really important,” he explains. “But it takes enormous computational power to be able to do that.”
We’ve got enough data – sometimes too much data – but now computation is the bottleneck.
That’s where the advanced technology comes in. A recent study by researchers at the Institute for Bioengineering of Catalonia, Barcelona, found that using machine learning over conventional computational methods reduced microscope data processing time from months to mere seconds.
Cambridge-1 also has the potential to drastically accelerate and optimise the later stages of drug research, enabling large volumes of valuable pathology and genomics data to be processed, enhancing researchers’ ability to generate genetically validated targets for drug candidates.
“If we just had two whole genomes, and we’re looking at a drug that we’re trying to use for a particular condition, then it’s not a very good representation of the whole world,” Rhodes says. “It won’t give us a very accurate understanding of, if we take this particular drug with this particular human with this type of genetic data, how that human might interact with that drug.
“We’ve seen this an awful lot; we saw it with the Covid-19 vaccine, where certain vaccines work very, very well on certain demographics, and some just didn’t,” he explains. “Understanding the genetic makeup allows us to understand some of the subtleties in those differences – but again, computationally doing that is massively challenging.”
One human genome equals around one gigabyte – roughly the size of a short film, Rhodes says. He compares processing large volumes of genomic data to having a computer review and analyse “every single scene, voice, motion, interaction” in thousands of Netflix movies all at once. Tasks like these – highly challenging for general-purpose computing systems – are a breeze for supercomputers like Cambridge-1.
Because the processing capabilities of supercomputers make it possible for scientists to process millions more samples than a traditional computing system could handle, diverse populations will be better represented, and researchers can more accurately determine which drugs are likely to be successful.
This saves drug companies both time and money; the computer’s data-based recommendations will allow researchers to create clinical trial cohorts that are more diverse, more precise, and therefore more likely to respond well to the drug candidate.
Improving diagnosis and treatment
Two of NVIDIA’s partners, King’s College London and Guy’s and St Thomas’ NHS Foundation Trust, are leveraging Cambridge-1 to “teach AI models to generate synthetic brain images by learning from tens of thousands of MRI brain scans, from various ages and diseases”.
It’s hoped the AI-based data model, which can generate an infinite amount of never-seen brain images with chosen characteristics, will allow a better understanding of brain diseases and enable earlier diagnosis and treatment.
Enhanced imaging processing is valuable in a number of therapeutic areas. Oncology, Rhodes says, is an area that can benefit hugely from NVIDIA’s technology.
“Let’s say we’re trying to cure a particular type of lung cancer. There are many types of lung cancers; we know some lung cancers, and we don’t know a lot about other lung cancers,” he explains. “So, people still die because we don’t know what that cancer was.
“We want to be looking at hundreds of thousands of [pathology] images of lung cancer, so that system has to look through hundreds of thousands of images and identify the variations in them – what’s the same? What’s different? What seems normal? What seems abnormal?”
That system has to look through hundreds of thousands of images and identify the variations in them.
According to Rhodes, these are simple activities, but for traditional computers the task of processing pathology images is laborious and slow.
“If we then say we want to look at the genetic data for this drug, for lung cancer, we need to look at hundreds of thousands of specimens,” he adds. “So that we can identify that this group of patients, with this particular type of lung cancer, have this particular trait within the DNA signature.
“If they have something that’s very specific, does this mean this drug is going to work well with that? Or does this mean that it won’t work well, or won’t work at all? And now, what’s becoming more advanced is to actually look at the image and then predict on the image where to look in the DNA data.”
While a pathologist can identify cancerous cells on an image within seconds, what isn’t always obvious is the kind of cancer being shown, or what stage the disease has progressed to.
“But if we can go into the DNA sequence data of that patient and identify where the particular markers we’re expecting for these types of cancer are, we can then go, ‘actually, we can see there’s a genetic issue here, or this genetic makeup would determine that it’s this particular type of cancer’,” Rhodes explains.
Even if analysis of a patient’s DNA doesn’t return any relevant or useful information, he says, this result could indicate that researchers are dealing with a type of cancer they’ve never seen before. This finding, too, would be valuable for investigating or assessing which treatment protocols may or may not work for a particular patient.
“The activities that we’re doing at this level, they seem very, very simple – it’s watching, it’s looking – but they’re hugely important to the outcome, whether it’s a patient or whether it’s a drug going through a particular process,” Rhodes says.
Is supercomputing the future of research?
Cambridge-1 isn’t the only supercomputer being applied to medical research. This year, the University of Nottingham’s School of Chemistry was granted access to Marconi100, a supercomputer developed by IBM and NVIDIA and based in Italy.
The professor leading the project, Jonathan Hirst, said it will “harness the power of both machine learning and of physics-based molecular simulation to accelerate the discovery of compounds with predicted therapeutic values as leads in drug discovery efforts”.
Elsewhere, Japanese scientists are using the world’s most powerful supercomputer, Fugaku, to accelerate the discovery and development of personalised therapies. Despite only officially launching in March this year, Fugaku has already contributed to efforts to identify existing medicines that can be used to fight Covid-19, narrowing over 2000 potential drug candidates down to just a few dozen.
We want people to realise that computational blockages, or challenges with computation, shouldn’t affect scientific research.
Back in 2019, eight sites across Europe were selected to host supercomputing centres aimed at supporting development in personalised medicine and drug and material design, among other applications. The centres were allocated by the European High-Performance Computing Joint Undertaking, which was set up to equip the EU with “a world-class supercomputing infrastructure by the end of 2020”.
If supercomputers are to transform medical research for good, the high-performance computing needed to truly optimise drug discovery and development must be widely available. Democratising advanced technology across the industry is crucial to discovering the most effective medicines, faster – and high-performance computers like Cambridge-1, designed specifically to aid researchers in their efforts, are a big step in the right direction.
“We want people to realise that computational blockages, or challenges with computation, shouldn’t affect scientific research,” Rhodes says. “We’ve done this to show the UK government, look at what could be done, and also look at the enormously valuable assets that are now being produced in the UK, by UK companies.
“How will [supercomputing] fundamentally change clinical research in the future? I don’t know the answer to that, but I know we’re giving them the assets to really make significant change,” he says. “Looking back on my life, I will really have thought that we have affected clinical change, we have improved patient lives, we have improved the NHS here – and NVIDIA’s been at the heart of it.”