How AI Steered doctors toward a possible Covid-19 treatment
How AI Steered doctors toward a possible Covid-19 treatment
Specialists at the London start-up BenevolentAI helped identify the arthritis drug baricitinib, which is now part of a clinical trial
By Cade Metz
Published: May 2, 2020
Specialists at the London start-up BenevolentAI helped identify the arthritis drug baricitinib, which is now part of a clinical trial. (Scott Gelber/The New York Times)
In late January, researchers at BenevolentAI, an artificial intelligence startup in central London, turned their attention to the coronavirus.
Within two days, using technologies that can scour scientific literature related to the virus, they pinpointed a possible treatment with speed that surprised both the company that makes the drug and many doctors who had spent years exploring its effect on other viruses.
Called baricitinib, the drug was designed to treat rheumatoid arthritis. Although many questions hang over its potential use as a coronavirus treatment, it will soon be tested in an accelerated clinical trial with the National Institutes of Health. It is also being studied in Canada, Italy and other countries.
The specialists at BenevolentAI are among many AI researchers and data scientists around the world who have turned their attention to the coronavirus, hoping they can accelerate efforts to understand how it is spreading, treat people who have it and find a vaccine.
Before the pandemic, the AI researchers were part of one of the most hyped and well-funded sectors of the tech industry, pursuing visions of autonomous vehicles and machines that can learn by themselves. Now they are simply trying to be helpful — working on technology that augments human experts instead of replacing them.
Medical researchers had spent years exploring baricitinib and similar medications as a way to treat viruses. Baricitinib, a pill taken once a day, can help fight extreme and unwanted activity from the body’s immune system, which occurs with both rheumatoid arthritis and viruses like HIV and can damage healthy cells and tissues.
In late January, after talking with one of the company’s investors in Asia about the pandemic, Baroness Joanna Shields, the chief executive of BenevolentAI, asked Peter Richardson, BenevolentAI’s vice president of pharmacology, if the company could explore potential treatments.
In an undated handout photo, clockwise from top left, Dr. Raymond Schinazi and Dr. Vincent Marconi, both of Emory University, and Peter Richardson and Olly Oechsle, both of BenevolentAI. Specialists at the London start-up BenevolentAI helped identify the arthritis drug baricitinib, which is now part of a clinical trial. (Handout via The New York Times)
BenevolentAI quickly joined a race to identify drugs that can block the virus from entering the body’s cells. Researchers at the University of California, San Francisco, and many others labs are looking into similar treatments.
BenevolentAI, which has received more than $292 million from the Singapore sovereign wealth fund Temasek, Goldman Sachs and others, had spent the past several years building technology that could help find information buried in vast troves of academic papers and other scientific literature.
The technology was designed for the development of new drugs — not for identifying new uses for existing medications — and it had never been used with material related to viruses.
Over two days, a small team used the company’s tools to plumb millions of scientific documents in search of information related to the virus. The tools relied on one of the newest developments in artificial intelligence: “universal language models” that can teach themselves to understand written and spoken language by analyzing thousands of old books, Wikipedia articles and other digital text.
These AI systems are rapidly improving everything from the Google Search engine to automated “chatbots” designed to carry on a conversation. They can also help machines comb through scientific literature, identify particular pieces of information, organize it and retrieve it on command.
Using its automated language tools, the company’s engineers generated a detailed and intricately interconnected database of particular biological processes related to the coronavirus. Then Richardson, who is 65 and a trained pharmacologist, used additional tools to browse through what the technology had found and understand what it meant.
“It is not like we have this giant button, and we just smack it, and stuff comes out the other end,” said Olly Oechsle, 37, the software engineer who oversees the design of these tools. “Peter has been working in this area since before I was born.”
Drawing on what the technology found in the literature, Richardson could map out the connections between particular human genes and the biological processes affected by the coronavirus. As a multicolored map appeared on his computer screen, two genes leapt out at him.
“They stood up and said, ‘Look, we’re here,’” Richardson said.
Once the genes were identified, he and his colleagues could pinpoint the way that existing medications targeted the genes, visualizing the process through a kind of digital flow chart. They identified baricitinib, made by the American pharmaceutical giant Eli Lilly.
Many scientists were already considering similar anti-inflammatory drugs that could reduce a cytokine storm, an extreme response from the body’s immune system that can kill coronavirus patients.
But the BenevolentAI researchers went further. Through their software, they found that baricitinib might also prevent the viral infection itself, blocking the way it enters cells. The company said it had no expectations for making money from the research and had no prior relationship with Eli Lilly.
Through Justin Stebbing, a professor of oncology at Imperial College London, the researchers sent their findings to The Lancet, one of Britain’s oldest and most respected medical journals, in early February. Like many other companies and researchers now exploring treatments across the globe, the team wanted to share what it had learned as widely as possible.
The next day, at Emory University Hospital in Atlanta, Dr. Vincent Marconi opened an email from a colleague, Dr. Raymond Schinazi, that pointed him and other colleagues to the paper. They had spent eight years exploring baricitinib and other drugs as a treatment for HIV, and they knew such drugs could potentially help coronavirus patients.
But they had not settled on baricitinib as a viable option, and they had not identified the specific properties that might allow the drug to fight the virus. Nor had the scientists at Eli Lilly.
At Emory, the lab researchers were shocked that the paper had come from BenevolentAI. “It was crazy,” said Christina Gavegnano, who took part in the work with HIV. “We kept asking, ‘Who are these people? Does anyone know them?’”
A month later, Marconi proposed a clinical trial with baricitinib and another drug. As coronavirus cases mounted at his hospital, he and his clinicians administered the pill as a compassionate measure to patients, with encouraging results.
“We normally talk about ‘bench to bedside,’” Stebbing said, referring to moving quickly from laboratory bench research to the treatment of patients. “This is about ‘computer to bench to bedside.’”
Unaware of what was happening in Atlanta, Mario Corbellino administered the drug as a compassionate measure at a hospital in Milan after reviewing the research from BenevolentAI and soon proposed another clinical trial. He and other infectious-disease specialists, he said, feel more comfortable testing this kind of drug if it has the potential to not just reduce an immune system response but prevent the viral infection.
Dr. Dan Skovronsky, chief scientific officer at Eli Lilly, warned that it was still unclear what effect the drug would have on coronavirus patients. Even after the clinical trial, he said, it may not be clear whether the antiviral properties pinpointed by BenevolentAI are as effective as they might seem to be.
He also said those properties were not something his scientists would have discovered so quickly on their own. “There is so much complexity to biology and there is so much information out there, it is hard — if not impossible — for one person to put together the clues that are already there in the literature,” he said.