AI Predicts Side Effects of Drug Interactions

Computer scientists from Stanford University have developed a convolutional neural network that can predict side effects when drugs are taken in combination.

The research effort could have a great deal of practical significance for patients, who increasingly are taking multiple drugs, either to treat a single disease or to alleviate a number of different health conditions. The problem is that interactions between drugs are often poorly understood and can carry significant risk.

According to Marinka Zitnik, a postdoctoral fellow in computer science, the problem is exacerbated by the sheer numbers of drugs available to doctors, noting “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.”

The problem led Zitnik and fellow researchers Monica Agrawal and Jure Leskovec to develop Decagon, an artificial intelligence application for predicting the potential side effects of different drug combinations.  To do this, they built a massive convolutional neural network that models how the more than 19,000 proteins in the human body interact with one another and how different drugs interact with these proteins. To drive this model, the researchers used databases that documented single protein-protein and drug-protein interactions. They also pulled in additional databases that detailed the side effects associated with certain drugs and drug combinations.

To get the level of predictive capability they were going after, the researchers opted to use a graph convolutional neural network.  These are commonly used for social networks and knowledge graphs, but they have yet to be adopted for computational biology. The model used by Decagon extends this design by incorporating support for multiple edge types, with each type representing a different side effect, and by implementing a unique form of weight sharing.

Decagon’s predictive capability is not perfect, but in many cases it turns out to be surprisingly accurate. In one case, the system suggested that muscle inflammation could be a possible side effect of combining atorvastatin, a cholesterol drug, and amlopidine, a blood pressure medication. Although none of the raw data used to develop the model suggested such an effect, according to a published case study, the combination did indeed lead to this condition, requiring one patient to be treated in an intensive care unit for 69 days.

After searching the medical literature for similar cases, the researchers found evidence for five of Decadon’s ten highest ranked predictions. (That doesn’t mean the other five are not valid, just that no documented examples have emerged.) Its predictive capability was especially good at finding side effects based on molecular interactions, but it was less accurate when side effects were based on the patient’s environment or behavior. In general, the researchers found that Decadon outperformed previous attempts at computational-based predictions by a large margin, writing that “across 964 side effect types, Decagon outperforms alternative approaches by 19.7% (AUROC), 22.0% (AUPRC) and 36.3% (AP@50).”

In its current form, Decagon only predicts side effects associated with pairs of drugs, but the team hopes to enhance the software to deal with more complex combinations. In anticipation of its use in a clinical setting, they also want to create a more user-friendly tool for doctors and medical researchers.

For those interested in the inner workings of Decagon and its design, you can read the research paper here.

Current rating: 5


There are currently no comments

New Comment


required (not published)