Using a new mathematical approach, scientists have predicted drug side effects that typically aren’t discovered until thousands of people have taken a medication. The technique is especially good at foreseeing side effects that show up after days or months of taking a drug, suggesting that a similar approach could help make drugs safer before they come to market and may even save lives.
Researchers started with a 2005 catalog of existing medications and their known side effects, such as heart attacks or sleeping problems. After linking drugs and their side effects into a network, they instructed a computer to predict likely new connections between drugs and side effects. The program was able to predict 42 percent of the drug–side effect relationships that were later found in patients, the researchers report in the Dec. 21 Science Translational Medicine.
“Adverse drug events are very important and understudied,” says Russ Altman, a biomedical informatics specialist at Stanford University who wasn’t involved with the work. Before a drug ever gets to market it undergoes toxicology testing and clinical trials to establish that it is effective but not dangerous. These trials are often extensive enough to prove that the drug works, but not big enough to say anything meaningful about side effects, says Altman. So, many side effects aren’t discovered until after the drug is on the market.
“You routinely find a whole bunch of annoying ones and every now and then there’s a showstopper,” Altman says. Such interactions lead to 770,000 injuries and deaths each year.
To clear some of the haze surrounding side effects, scientists from Harvard Medical School and Children’s Hospital Boston created a network linking 809 medications to 852 side effects that were known as of 2005. The team also added information to their network on chemical properties, such as the drug’s melting point and molecular weight, and where the drug does its stuff in the body. Using these data and relationships alone, the computer predicted side effects that were reported in later years, such as the seizure drug zonisamide causing suicidal thoughts in some people and the antibiotic norfloxacin’s link to ruptured tendons. It also linked the controversial diabetes drug Avandia (rosiglitazone) to heart attacks, a connection that is supported by some research.
The team tried adding additional information about drugs, such as data describing molecular structure. But the network diagram of the known relationships between drugs and side effects alone had more predictive power and fewer false positives than methods that added the additional information, the team reports.
“We were pleasantly surprised,” says team member Ben Reis, who directs the predictive medicine group at Children’s Hospital Boston. “The network encodes a lot of information from other worlds. Perhaps that’s why it did so well.”
There were some side effects for which the model performed less well, such as skin problems, notes mathematician Aurel Cami of Harvard Medical School and Children’s Hospital Boston.
This first round established that network math, typically used for assessing social relationships or how a disease spreads, can uncover important drug reactions. Now Reis and Cami are investigating what kinds of data work best and trying to tackle drug-drug interactions that can also be dangerous and are rarely studied in clinical trials.
“We’re moving from a paradigm of detection — where it takes sick people to know something is wrong — to prediction,” says Reis.