The artificial intelligence start-up Insilico Medicine has used machine learning to find credible drug candidates in a matter of weeks (Nature Biotechnology 2019, DOI: 10.1038/s41587-019-0224-x). Experts say it’s an important demonstration of what machine learning can do in drug discovery, but it isn’t a revolution.
Insilico Medicine has been showing off its progress in teaching computers to find new drugs since its founding in 2014. The company’s latest effort involves generative reinforcement learning, a technique that uses rewards to guide an algorithm as it searches for molecules that satisfy its goals. In this case, the algorithm was hunting for small molecules that are inhibitors of discoidin domain receptor 1 (DDR1), a kinase that is linked to fibrosis.
The researchers trained their algorithm using databases of known DDR1 inhibitors, kinase inhibitors, nonkinase inhibitors, and patent-protected molecules. Based on measures of novelty and DDR1 inhibition, the algorithm proposed 30,000 potential drugs. This group was filtered by computer programs and by chemists to a final six candidates. Two of them showed no activity, but the researchers synthesized two others that showed promise. One compound performed well against kinase screens, and the researchers tested its metabolic stability in mice. The whole process took 46 days.
Some chemists familiar with drug discovery and AI applaud the effort. “This is an impressive demonstration of rapid hit expansion starting from a pool of known kinase inhibitors,” says Connor Coley, a computational chemist at the Broad Institute of MIT and Harvard. But he and others agree that the caveat is in the second half of that sentence. The molecules that the algorithm identified look similar to other kinase inhibitors, and Insilico Medicine had a lot of public information about similar compounds to train it on, notes Ingo Hartung, director of medicinal chemistry at Merck KGaA. Hartung likens what the algorithm did to using Google Maps to find New York City’s Time Square: It did a good job, but the job wasn’t a hard one. Hartung says the process’s speed is impressive, but he wants to see it tested on harder problems.
Other machine-learning experts are less convinced that Insilico Medicine’s 46-day time line is such an achievement. Traditional drug-discovery techniques might have worked just as quickly, says Olexandr Isayev, a computational chemist at the University of North Carolina at Chapel Hill. The researchers don’t provide a baseline for comparison. Without that, adds Ash Jogalekar, a medicinal chemist at the AI-oriented biotech firm Strateos, “it’s thus impossible to know whether the results attributed to the technique are unique in any way or not.”
Alex Zhavoronkov, Insilico Medicine’s founder and CEO, says the group has already tested the method on more challenging problems but hasn’t made the results public. Zhavoronkov also says the hits found in this research are being tested in disease models other than fibrosis.
He says the current work is a small piece of what needs to be done to make drug discovery by artificial intelligence successful. Identifying the molecules is important, Zhavoronkov says, but molecules are worthless without validation. That’s what Insilico Medicine plans to do next.