Stanford AI can predict negative side effects of millions of drug combinations

Nearly 40 percent of Americans over the age of 65 take five or more different drugs, and doctors often simply have to monitor patients to see if any of those drugs combine to create adverse side effects. 

Drug combinations are a remarkably unstudied area, but as Marinka Zitnik explains, "it's practically impossible to test a new drug in combination with all other drugs, because just for one drug that would be five thousand new experiments."

So Zitnik and her Stanford colleagues set out to find a solution to the problem. They created a massive deep learning system trained on data encompassing over 19,000 proteins and how different drugs interact with those proteins. The system is called Decagon, and it can effectively predict the consequences of combining any two different drugs.

To test out Decagon's predictive abilities the team examined 10 of the system's predicted drug pair interactions that didn't have clearly known adverse interactions. The researchers found new supporting case study evidence backing up five of those 10 predictions.

For example, one prediction from Decagon suggested muscle inflammation would be caused by combining a certain cholesterol drug with a blood pressure medication. This adverse side effect of combining those two drugs was only recently verified by a case study published in 2017.

The next stage in the project is to try to turn Decagon into a more user-friendly tool that doctors can easily navigate for information when prescribing combinations of drugs. At this stage, the system only evaluates drug pairs but the researchers hope to expand that into more complex combinations of drugs in the future.