From aerospace to mechanical engineering, professionals across the world rely on powerful supercomputers to aid their work each day. But in the pharmaceutical industry, one of the most high-tech industries in the world, the majority of drug development is still done by hand.
Researchers physically synthesize each and every promising compound, then test to see if it’s safe for human use. Creating a drug that effectively treats an illness while producing the least amount of side effects is a process that can take years—and the cost of research has been rising dramatically, too.
Videos by VICE
But a young Canadian startup called Atomwise thinks it has an alternative approach. The company, which was started by a computer scientists at the University of Toronto’s Impact Centre, has created a machine learning algorithm that they hope will not only help researchers develop the next generation of pharmaceutical drugs, but do so faster and cheaper than ever before.
According to a report released in 2003 by the Tufts Center for Drug Development, pharmaceutical companies were paying about $802 million ($1.04 billion when adjusted for inflation) to develop a new drug between 1983 and 1994. When Tufts completed a follow up study in 2013 that looked at the cost of drug development between 1995 to 2007, development costs had risen to $2.5 billion per successful drug.
“There are health threats out there for which there are no adequate solutions today, and it’s unclear if the status quo will be enough”
The increase in cost is due to several factors, the most important of which is that, as scientist attempt to develop treatments for more complicated chronic and degenerative diseases, more and more drugs are failing to pass human trials. It’s currently estimated that eight out of 10 drug development projects are abandoned.
However, there are other factors, as well. According to the Tufts report, increased trial complexity and larger clinical trial sizes have all led to rising development costs.
It’s a trend that, if allowed to continue, will have worrisome consequences on human health. “There are health threats out there for which there are no adequate solutions today, and it’s unclear if the status quo will be enough,” said Alexander Levy, chief operating officer of Atomwise.
Levy says one of the health issues he and his colleagues are most worried about is antibiotic resistance. Antibiotics that were developed in the 40s, 50s and 60s are becoming increasingly ineffective in treating bacterial infections— and he’s not the only one that’s worried. According to the CDC, each year some 23,000 Americans die from bacteria that have become resistant to current antibiotics.
“One day even more recent antibiotics won’t work either,” he said. “That will lead to a world where people can’t have safe routine surgery and a world where minor injuries may lead to fatal infections. We will live in a world where mortality goes dramatically up.”
It’s not just bacterial infections that could pose a serious threat to human health, either. A host of other illnesses—including the resurgence of diseases like polio, measles and whooping cough, as well new strains of the avian and swine flu—all have the potential to severely stress the scientists’ ability to develop new treatments fast enough to keep up.
The algorithm Atomwise developed is similar to the Deep Learning Neural Networks used by DeepMind, a startup that was acquired by Google last year for $628 million. While Google has been happy to let the AI teach itself how to play Space Invaders, Atomwise has asked it to learn complex biochemical principles instead.
“We let it take thousands of simulated [processor] years to teach itself the factors that are ultimately most predictive when it comes to the effectiveness of a drug,” said Levy. “Our system doesn’t look at a dozen or two dozen factors, it looks at thousands of factors at the same time and combines them in complicated and nonlinear ways.”
According to Levy, this is what makes his company’s system unique. He says it’s an approach that is similar to how computers go about image recognition.
The way in which humans differentiate between an image of a cat and a dog is vastly different than that of a computer program. A human will intuite the difference, but a computer will look at factors like stripes and colouring—factors that, at first blush, seem unintuitive—to differentiate between the two. In the same way, Levy says their system has devised some unintuitive methods for discerning what chemicals will properly latch onto a biological target.
But Atomwise’s main challenge now is to convince big pharma that supercomputers are the panacea to its financial woes.
According to Dr. Vijay Pande, a professor at Stanford University and the director of the folding@home, a project that runs protein folding simulations on thousands of computers across the world, pharmaceutical companies have good reasons to be reticent about introducing supercomputers into their practice. “A lot of them have been burned,” he said. “In the 1980s, a lot of people promised that computers would revolutionize drug design. It really was oversold and overhyped, there’s little doubt about that.”
However, he says that things are slowly starting to change, in part thanks to projects like folding@home that are helping to show the industry how computers can assist with drug research.
Dr. David Topham, executive director of the University of Rochester’s Health Sciences Center for Computational Innovation (HSCCI), said he and his colleagues initially “had a lot of skeptics, asking us why we needed so many big computers,” he said. The medical research facility is now home to 420 teraFLOPS of computing power, and its supercomputers are always at capacity.
“People were not demanding that we install supercomputers, but we did so anyway,” Dr. Topham said. “Now we’ve created a community that can’t imagine being without this technology.”
Globavir, a company that Dr. Pande co-founded, is already using supercomputers to assist in its own drug development. Drugs that the company developed using software algorithms to treat dengue fever—a mosquito-borne disease that is common in the tropics—and the West Nile virus are currently in clinical trials. Like Atomwise, the company claims that its drug discovery platform is able to significantly reduce the time it takes to develop a new pharmaceutical treatment.
Elsewhere, the United States Department of Energy, through its Oak Ridge Leadership Computing Facility, has also developed a supercomputer assisted drug discovery system.The scientists at Oak Ridge claim that their system can simulate how two million different types of drug compounds interact with a specific target in less than two days.
Where these systems live and die, says Levy, is in the accuracy of the results they produce. Atomwise’s software currently runs off of an IBM Blue Gene/Q with a peak performance of 838 teraFLOPS. This hardware enables the company’s software to identify the most promising drug candidates in a matter of weeks, not years, and Levy claims the system he and his colleagues have developed is as much as 18 times more accurate than previously used methods.
“If this technology is adopted, it will dramatically change the economics of drug development,” said Levy. “It will help address the long standing problems of orphan and neglected diseases, and finally it will actually help us tackle what are likely the biggest global threats that we’re going to face as a species in the next couple of decades—threats so large we may have no adequate methods other than this.”