Startup uses A.I. to identify molecules that could fight coronavirus

February 6, 2020, 2:00 PM UTC

Insilico Medicine, a startup based in Rockville, Md., says it has used artificial intelligence to rapidly identify molecules that could form the basis of an effective treatment against the coronavirus at the heart of the current outbreak.

It took Insilico’s A.I.-based system four days to identify thousands of new molecules that could be turned into potential medicines against the virus. Insilico says it will synthesize and test 100 of the most promising candidates, while publishing the full library of new molecular structures it has generated for other researchers to possibly use.

The global health emergency over the spread of the deadly new coronavirus, known as 2019-nCoV, is proving to be a crucial real-life test case for a host of new biomedical technologies, as well as new organizations and funding bodies, that aim to dramatically reduce the time it takes to create new vaccines and drugs to combat emerging pandemics.

The idea is to find and test new treatments, putting them into human clinical trials in as little as a week or two, as opposed to the years it might otherwise take. U.S.-based biotechnology firm Gilead struck a deal earlier this week with a Beijing hospital to begin immediate human testing of an existing antiviral drug, remdesivir, in Wuhan, the heart of the coronavirus outbreak.

Insilico only decided to see if it could come up with interesting leads for a new 2019-nCoV treatment on Jan. 28, Alex Zhavoronkov, the company’s founder and chief executive officer, says. “When the virus outbreak started, we didn’t realize it was serious,” he says.

Once it decided to get involved, Zhavoronkov says, the company studied the long list of possible targets for a 2019-nCoV treatment that have been published by the Beijing-based Global Health Drug Discovery Institute. It decided to go after an enzyme, called 3C-like protease, that is critical for the coronavirus’s reproduction.

Zhavoronkov says Insilico picked this target, in part, because it is similar to other viral proteases whose structures were previously mapped and because it had access to a model of the 2019-nCoV 3C-like protease itself that had been developed by Zihe Rao, a well-known expert on virus protein structures at Shanghai Tech University.

Beginning on Jan. 31, Insilico began using 28 different machine learning models to design new small molecules that might bind to the 3C-like protease and inhibit its functioning.

Some of these techniques employ generative adversarial networks (or GANs), the same kind of machine learning that is best known for creating deepfakes. In this case, instead of producing highly-realistic, fake videos, the A.I. generates new molecules that will form the right structure to bind with the protease.

Insilico uses further machine learning techniques to filter the molecule designs these GANs produce: it favors those that score highly for “drug-like” properties and for being chemically active, while tossing out molecules that, due to their properties, are unlikely to ever work as drugs, such as metal compounds.

It also applies filters to make sure the overall set of molecules it is generating are unlike existing, known structures (Zhavoronkov says that no molecule its system creates shares more than 70% of its structure with a molecule already discovered), and are diverse from one another, so that the company has a good range of candidates to test.

Four days later, Insilico’s software had generated hundreds of thousands of new molecule designs, and filtered them down into thousands that met its criteria for possible drug candidates. “In four days we generated pretty good molecules,” Zhavoronkov says.

The company published a paper detailing its research on the free, non-peer reviewed research repository Research Gate. The company also published its research and the designs for all of its potentially-useful molecules on its own website. It is calling on researchers to examine and critique the molecules its system has generated in the hopes of accelerating the process of finding ones that could be useful for a coronavirus treatment.

The company is not the only one hoping that A.I. will help create new treatments for the Wuhan coronavirus. A team from Michigan State University last week also published a paper on using machine learning techniques to create new drug candidates for 2019-nCoV.

Insilico, which was founded in 2014 and has raised about $50 million in venture capital funding to date, uses a number of different A.I. techniques to invent new molecules that can form the basis of pharmaceuticals, determine if existing drugs can be repurposed for new uses, and to predict the outcome of clinical trials. The company has a partnerships with pharmaceutical giant GSK and China’s Jiangsu Chia Tai Fenghai Pharmaceutical to help those companies find molecules for possible new drugs.

This story has been updated to reflect new information from Insilico Medicine about where its research has been published and to include updated links to that research. Insilico originally said it published on the research repository Biorxiv. It published on Research Gate instead.

More must-read stories from Fortune:

—The first coronavirus drug candidate is set for testing in China
—The Gates Foundation is pledging up to $100 million to fight coronavirus
—World’s biggest brewers tally costs of coronavirus outbreak as beer sales suffer
—Governments deploy surveillance tech to track coronavirus victims
Asia worries about big events like the Olympics amid coronavirus spread

Subscribe to Fortune’s Brainstorm Health for daily updates on biopharma and health care.

Read More

Artificial IntelligenceCryptocurrencyMetaverseCybersecurityTech Forward