Wielding AI to discover druglike molecules at speed and scale
CEO: Alex Zhavoronkov, Ph.D.
President: Alex Aliper
Based: Hong Kong and Rockville, Maryland
The scoop: We’ve been told for years that artificial intelligence could one day be the key to faster drug discovery, development and testing—an essential tool to shorten the time between an idea and a potential cure. Last fall, we started to see some proof.
In fewer than 50 days, Insilico Medicine demonstrated it could use AI to create thousands of druglike molecular structures and winnow them down to a promising few, before successfully testing the lead candidate in live mouse models.
The company set its AI’s sights on drugging a receptor tyrosine kinase known as DDR1—found on epithelial cells in the skin, kidney, liver and lungs—that is linked to diseases and disorders tied to fibrosis.
In three weeks, Insilico had developed a slate of 30,000 novel, computer-generated small molecules. In 25 days more, it selected and synthesized the most promising six and ran them through in vitro testing for selectivity and metabolic stability before an in vivo experiment.
The paper was published in Nature Biotechnology last September and has since become one of the journal’s most popular papers. Additionally, the code used for the project—dubbed GENTRL, for generative tensorial reinforcement learning—was made open source and available to the public.
Now, the company is turning its attention toward finding a therapeutic for the novel coronavirus—and because it’s not looking to disclose all its methods in a journal, Insilico can throw its full, proprietary pipeline behind the effort.
“In four days, we managed to generate 100 druglike structures,” says co-founder and CEO Alex Zhavoronkov, Ph.D., who described using separate AI networks based on which pieces of the coronavirus’ molecular puzzle were already known. In early February, the company posted those molecules online to be evaluated by medicinal chemists and other researchers.
“We actually managed to get the entire community to contribute something and comment,” Zhavoronkov describes, with some volunteers running docking simulations testing the molecules against the virus’ key proteins. “So now we have all those comments, and we start synthesizing.”
Insilico expects it will take four to six weeks to deliver in vitro testing results, due to their complexity, with a compound potentially ready for human testing later this year. “These aren’t known drugs—there’s no known chemical routes, there is just absolutely no [previously established] chemistry,” he says.
Zhavoronkov never thought the company would take a deep dive into antibiotic or antiviral drug development, a challenging area that has been consistently underfunded. “I would say it’s not only not a growing field, it’s a cemetery,” he says.
But now—with the coronavirus providing a real-world impetus to demonstrate the speed and cost-effective benefits of AI-generated molecules—it's flirting with the idea. Spinoffs, joint ventures and externalized projects are nothing new to the company, and Insilico has been prolific: Zhavoronkov counts about 30 active endeavors borne out of the company’s programs, ranging from breast cancer to Duchenne muscular dystrophy to a recent foray into brain cancer.
“This case even affected our company’s strategy a little bit, because we might go the anti-infective route as well,” he says.
What makes Insilico Medicine fierce: The company’s culture is built on competition, from within and without. In the digital space, it’s employed generative adversarial networks: pairs of machine learning systems that jockey with one another to produce and evaluate better molecular candidates.
Elsewhere, it uses hackathons and international coding competitions to select and hire its researchers, and it grants company rings to those who reach certain levels in its management.
“We started hiring through competitions in 2015, and we hired some of the brightest minds in the country at that time,” including recent Ph.D.s in AI and machine learning, Zhavoronkov says. “None of them knew anything about biology, … so we had to spend three years to educate them on biology.”
That illustrates the position where Insilico resides, in something of a valley between the high-tech and Big Pharma camps. To bridge the two, the company has worked to convince them it belongs in both.
“One of the challenges we had in AI is that when you are submitting a high-end algorithm to an AI conference, and you show them that you did the work on molecules, they will not accept the paper,” Zhavoronkov says. “For three years we struggled.”
“They told us to submit to a more specialized journal, or said they didn’t understand the chemistry,” he says. “Last year, we decided to submit work on pictures—we took people’s faces and did this deep-fake thing—and immediately we got a paper accepted. And it was actually very similar to what we had submitted previously, but in molecules.”
On the other side, many expert medicinal chemists don’t trust AI until they’re shown strong, working molecules that don’t contradict their past training and experience, he says.
“We’re trying to get our AI guys to learn medicinal chemistry, to learn biology, and they like it—well, some of them don’t, but they don’t have much choice,” Zhavoronkov says. “They need to understand the fundamentals and they need to understand what medicinal chemists want.
"At the end of the day, it’s all about the molecule," he adds. "Doesn’t matter how you make it.”