AI system predicts billions of potential side effects from drug combinations

An artificial intelligence system developed by researchers at Stanford University could ultimately help physicians make more informed decisions about which medications to safely prescribe.

The AI system, called Decagon, is able to predict billions of potential side effects from drug combinations—no small feat given that there are nearly 125 billion possible side effects between all possible pairs of drugs, most of which have never been prescribed together or systematically studied.

Researchers contend that the predictions generated by the deep learning approach, once provided to doctors in a more user-friendly format, would significantly improve the current capability available to clinicians—namely, mere chance.

With about 5,000 approved drugs on the market and 1,000 different known side effects, “it’s practically impossible to test a new drug in combination with all other drugs, because just for one drug that would be 5,000 new experiments,” says Marinka Zitnik, a postdoctoral fellow in computer science, who helped design the AI system with her colleagues at Stanford.

“People take multiple drugs at the same time, and we really don’t know much about what are the side effects associated with those drugs,” adds Zitnik, who notes that—according to the CDC—39 percent of Americans over age 65 take five or more prescription drugs. “We asked the question: is it possible to somehow computationally model drug combinations, not just individual drugs?”

The goal of the research was to identify safe and unsafe combinations of drugs before they are used by patients, Zitnik emphasizes.

p1aga522jul091havinc1h92cv0h.jpg
Stanford university campus in California

Also See: MIT machine learning makes brain cancer treatment less toxic

By leveraging more than 4 million known associations between drugs and side effects, researchers designed a deep learning method to identify patterns in how side effects emerge based on how drugs affect different proteins in the human body.

Jure Leskovec, an associate professor of computer science at Stanford, points out that there are more than 19,000 proteins in the human body that interact with each other and—in turn—different drugs affect these proteins in different ways.

“The key to our approach is this notion that proteins are these worker molecules in cells,” says Leskovec, who calls the neural network that researchers composed, describing how proteins in bodies work, as a “map of human life, if you will.”

“The way drugs work are by affecting the activities of the proteins,” adds Zitnik. “When a patient takes a drug, that drug targets certain proteins and changes their activity in the cell—such that, it hopefully addresses the problem of the disease.”

As an example of Decagon’s predictive powers, Zitnik recounted how the AI system correctly predicted that the combination of cholesterol drug atorvastatin and blood pressure medication amlopidine could lead to the serious side effect of muscle inflammation.

“Today, drug side effects are discovered essentially by accident,” added Leskovec, “and our approach has the potential to lead to more effective and safer healthcare.”

Currently, Decagon only analyses side effects associated with pairs of drugs. However, going forward, researchers want to tackle more complex regimens—including personalized information about patients, according to Zitnik.

“Not everyone that takes a particular combination of drugs will develop a particular side effect,” she concludes. “What would be great would be able to link the Decagon system with personalized medical records or physician notes for patients.”

“The other interesting direction for future work is to not only predict adverse side effects but also positive side effects,” comments Leskovec, who sees their deep learning approach ultimately helping to maximize the efficacy of medications and to identify better combinations of drugs to treat complex diseases.

For reprint and licensing requests for this article, click here.