Brand new drugs can take years to safely develop, but COVID-19 has prompted a rush to find therapeutics faster than ever.
Since March, Prof. Marinka Zitnik (Department of Biomedical Informatics) has applied graph neural networks, a type of algorithm, to systematically parse through existing drugs, searching for those that have therapeutic effects against COVID-19. She will speak about “Graph Neural Networks for Identifying COVID-19 Drug Repurposing Opportunities” at the AICures Drug Discovery Conference on October 30th.
Zitnik and her group rely on knowledge graphs—which link diverse types of data into vast networks—to map 6,340 drugs already on the market. Their method takes into account everything from chemical structures and molecular profiles, to known adverse effects, and data in electronic health records. Testing out several kinds of algorithms, the team generated a multiple ranked lists of the drugs, based on how likely they could have a therapeutic effect against COVID-19.
Throughout the process, the group followed ongoing critical trials. For any of the 6,340 drugs, being in a trial indicated that the medical community had put stock in that drug. It was a way to loosely test predictions, before those predictions were validated in a lab, Zitnik says, although it was important to keep some caveats in mind. Some drugs, hydroxychloroquine for example, entered more trials after being highly politicized or receiving media attention. Many COVID-19 trials are yet to be completed, making it necessary to find experimental data elsewhere.
“For us, it was really important to be able to work with domain experts and biologists to validate our predictions,” Zitnik says. “Then we went back and calculated the performance of the algorithms.”
Starting in April, the group collaborated with microbiologist Robert Davey and others from the National Emerging Infectious Diseases Laboratories (NEIDL) at Boston University to validate their ranked lists in a wet lab. Davey and his colleagues investigated 918 of the top-ranked drugs. To test for a possible therapeutic effect, researchers combined each drug with monkey kidney cells that had been infected with SARS-CoV-2. After observing the culture for several days, the group estimated the efficacy of each drug by how many cells survived infection. By comparing the experimental results with their initial predictions, Zitnik and her team, together with the task force of Prof. Albert-László Barabási from Northeastern and Prof. Joseph Loscalzo from Harvard, were able to refine the separate algorithms. They combined the improved algorithms with their initial work and experimental results to create a single prediction that outperformed the others.
Out of 918 drugs, 77 showed an effect against COVID-19. Notably, only one drug binds directly to proteins targeted by the virus. All others are considered network drugs, meaning they wouldn’t be identified with target-based drug discovery, which is commonly used.
“The observation is significant because it indicates how networks are important and powerful to consider when making those predictions,” Zitnik says. “The proteins and molecules within a cell are not independent of each other, they interact.”
Within the generated list, two drugs—an antihistamine for allergies and a heart failure drug—ranked high on the list, but aren’t in clinical trials for COVID-19. In other cases, drugs currently in clinical trials ranked very low. Zitnik sees several interpretations: the current data may be incomplete, accounting for an imperfect list. Or, “there’s a discordance between what’s being tested versus what actually works,” she says.
Recently, Zitnik has begun working on predictions for combinatorial therapies—combining drugs to achieve a synergistic, therapeutic effect against COVID-19. So far, the group has generated a large number of predictions, and are communicating with domain experts to prepare for experimental validation.
“It’s hard to expect that a single drug, which, of course, was initially not developed for Covid, to just magically have the desired properties,” she says.
In the future, Zitnik hopes to go beyond the binary of whether a drug works or not. She hopes to generate more nuanced predictions that account for biological factors.
“It’s not only about the drug, will it inhibit and kill the virus or not. There’s lots of subtle questions,” like what dose is needed, or how quickly a drug works, she says. “There’s lots more exciting work that needs to be done here.”