AI Is Screening Billions of Molecules for Coronavirus Treatments
AI Is Screening Billions of Molecules for Coronavirus Treatments

AI Is Screening Billions of Molecules for Coronavirus Treatments

Machine learning has pegged existing drugs to repurpose for COVID-19 clinical trials.

Abby Olena
Abby Olena
May 7, 2020


As COVID-19 cases continue to rise, physicians have begun to repurpose existing drugs in an attempt to find something that will help patients get better. In a pilot study at the end of March, 12 adults with moderate COVID-19 admitted to the hospital in either Alessandria or Prato, Italy, received a daily dose of the rheumatoid arthritis drug baricitinib, along with an anti-HIV drug combination of lopinavir and ritonavir, for two weeks. Another study group of 12 received just lopinavir and ritonavir.

After their two-week treatment, the patients who received baricitinib had mostly recovered. Their coughs and fevers were gone; they were no longer short of breath. Seven of the 12 had been discharged from the hospital. In contrast, the group who didn’t get baracitinib still had elevated temperatures, nine were coughing, and eight remained short of breath. Just one patient from the lopinavir-ritonavir–only group had been discharged.

The study comes with serious caveats: the lack of a formal control group, the small number of subjects, and the open-label design, meaning both physicians and patients knew which course of treatment they received. But it wouldn’t have been conducted at all if not for the work of BenevolentAI, an artificial intelligence (AI) company based in the United Kingdom. Researchers there, along with collaborator Justin Stebbing, an oncologist at Imperial College London, published a letter to The Lancet on February 4, describing how they used AI to identify baricitinib’s potential to treat COVID-19.

AI “makes higher-order correlations that a human wouldn’t be capable of making, even with all the time in the world. It links datasets that a human wouldn’t be able to link,” explains Stebbing.

Evaluating 1 billion small molecules for their ability to bind SARS-CoV-2 proteins would take a decade on even the largest of supercomputers.

The work also moved fast—crucial in a pandemic. Peter Richardson, BenevolentAI’s vice president of pharmacology, says that it took him only an afternoon of work to use the company’s knowledge graph—an enormous, digital storehouse of biomedical information and connections inferred and enhanced by machine learning—to identify two human protein targets to focus on, AP2-associated protein kinase 1 (AAK1) and cyclin g-associated kinase (GAK). These kinases mediate endocytosis—a process by which cells engulf things, including viruses—and, if disrupted, might make it harder for SARS-CoV-2, the virus that causes COVID-19, to get into human cells. Once the researchers had those targets, they used another algorithm to find existing drugs that could hit the protein targets. The team completed the work in only a few days.

By eliminating all the drugs that are not approved by regulators, the researchers cut the list down to about 30, and sorted out the handful of those that showed the highest affinity for binding their targets. Two were toxic chemotherapy drugs, and of the remaining three, baricitinib was the clear winner. In clinical trials as a treatment for rheumatoid arthritis, the side effects were mostly benign and showed up after a longer period of treatment than COVID-19 patients are likely to need. Plus, it’s not metabolized by the liver and is instead excreted through the kidneys, meaning that it might be safe to combine it with a traditional antiviral—such as lopinavir—that is metabolized by the liver.

In addition to baricitinib’s predicted interactions with AAK1 and GAK, it’s a known Janus kinase (JAK) inhibitor. Because JAK mediates cytokine signaling that leads to inflammation, inhibiting JAK suppresses inflammation, which, at first blush, might have been a problem.

“We were in this anomalous position of having a drug that was anti-inflammatory and saying we should give this to someone who is infected by virus, which doesn’t make a lot of sense,” says Richardson. But the idea is that baricitinib not only prevents the virus from getting into cells, but also reduces the intense immune reaction that causes so many problems, even as viral titers start to fall.

It was a good enough idea that Eli Lilly, the pharmaceutical company that makes baricitinib, has entered into an agreement with the National Institute of Allergy and Infectious Diseases to study the drug’s effectiveness in COVID-19 patients in the US.

“Even if the trial doesn’t work, we’re going to find out a huge amount of who it might work in and when it might work,” says Stebbing. “It’s all about personalized medicine, which means treating the right person at the right time with the right disease with the right drugs. Hopefully, this will be a powerful part of the jigsaw.”

The BenevolentAI team is one of several groups leveraging AI to find drugs that have already been approved by regulators and could therefore be repurposed to fight SARS-CoV-2.

Shantenu Jha, a computational scientist at Rutgers University and Brookhaven National Laboratory, is coupling artificial intelligence techniques and algorithms with high-performance computing simulations to speed up the ability to screen billions of existing drugs for their interactions with and ability to disrupt SARS-CoV-2 proteins. In a pandemic, “there are many drug candidates people would like to screen and, even with the proliferation of cloud and supercomputers, there just wouldn’t be enough computing” to test them all, he says.

Evaluating 1 billion small molecules for their ability to bind SARS-CoV-2 proteins would take a decade on even the largest of supercomputers, Jha says. But he and his colleagues have integrated machine learning into the simulations they run on the supercomputers, which allows the programs to adapt to new information that gets uncovered as they run, thereby producing results—in this case, lists of candidate small molecules—much faster than traditional supercomputing methods.

Each Friday, team members share a list of top candidates with collaborators, who either use their own AI-based methods to assess the small molecules virtually or test their effectiveness against coronavirus in the lab. Whatever the results from those analyses, the researchers feed them back into the simulations to refine the search for drugs.

“With AI, the whole idea is the more accurate data you can give it, the better its ability to predict, guide, whatever you’re using it for. You can never get enough good data,” says Jha. As the researchers tweak the simulations in response to community input, they aim to increase their chances of landing on a drug candidate that could make a difference in the pandemic.

Another group, led by Albert-László Barabási, a researcher at Northeastern University, is combining AI with another strategy, network medicine—the idea that by understanding the genetic and protein interactions in the body, researchers can gain a better understanding of how things might go wrong during disease. The team has a suite of tools that connect what’s known about these interactions and how existing drugs fit into the network. Then, the researchers pinpoint the aspects of the network, or neighborhoods, that are perturbed during disease and use AI and network associations to find existing drugs that could be repurposed to offset those disruptions.

One of the things that our methods don’t tell us is whether [these drugs] would be making the situation better or worse.

— Albert-László Barabási, Northeastern University

Sometimes, drugs target the proteins directly involved in a disease, Barabási explains, but more often they hit somewhere else in the neighborhood. Over the course of three weeks, he and his colleagues searched for interactions between approved drugs that could be repurposed and the network neighborhoods of the human proteins shown to bind to SARS-CoV-2 proteins. (Another group had identified these viral proteins in a March 22 bioRxiv preprint, which has now been published in Nature).

“The network medicine and AI tools have learned from what we know about the existing drugs and their network-based relationship . . . and now find similar patterns in the case of the drug and the COVID targets,” he says. “The hope is that if you find such a drug, the drug would either block the virus’s ability to do what it normally does or balance its consequences.” Perhaps the drug would interfere with a protein the virus needs to bind to as a means of infecting the cell or stop the coronavirus from using the host cell’s machinery to create its own proteins.

In an arXiv preprint submitted on April 15, the Barabási group shared a list of 81 approved drugs that the team and its collaborators are now testing in the lab to see whether or not the drugs interrupt a coronavirus infection and how they might do so. Some of the drugs—such as ritonavir, lopinavir, and chloroquine—are already being tested in COVID-19 clinical trials, while others, such as the antihistamine azelastine, are not.

“One of the things that our methods don’t tell us is whether [these drugs] would be making the situation better or worse,” says Barabási. Their strategy just indicates whether a drug perturbs the right neighborhood, not how it will affect the network once it does. “That’s why we prefer to start with the lab-based experiments before clinical trials because in the lab we will be able to see if the impact goes in the right direction.”

Bartosz Gryzbowski, a chemist at the Ulsan National Institute of Science and Technology in South Korea, and his group have used a small portion of their AI platform that is usually dedicated to novel drug development to analyze drugs and their interactions at the molecular level. In a ChemRxiv preprint posted April 17, the researchers used AI algorithms to identify progeny drugs that bear similarities to the drugs already being tested for COVID-19, which they termed parents.

The platform evaluates a parent drug’s three-dimensional structure, the specifics of its structure that lead to chemical interactions, and the context of how the overall structure and these specific domains interact with a potential target and with other domains within the molecule. Then, it can identify progeny drugs that have the potential to interact at a molecular level in a similar way. In the case where a parent drug proves to be ineffective in laboratory or clinical trials, having the progeny drugs already identified could point the way for clinicians who want to try alternatives quickly.

“It might be an act of desperation to go for drug repurposing, but also for this pandemic and for the next pandemic . . . there must be some fundamental shift in how we discover or how we at least get hints about drugs because every time some virus happens, we’ll be completely unprepared,” Gryzbowski says.