A major challenge in drug development is figuring out what might go wrong. During the development process, a new drug might be given to a few thousand people, maybe fewer if it’s for a rare or orphan disease – just enough to tell whether it does what the researchers think it will and to establish its short-term safety.
Once a drug is approved and available to the public, and out of the controlled laboratory or clinical trial environment, regulators rely on a mix of surveillance, reporting (by doctors and patients) and data mining to catch problems.
But these methods can fall short when it comes to rare side effects, drug-drug interactions or adverse events that arise only after patients have been on a drug for a long time. It can be years before doctors and regulators gather enough data and address safety problems with label warnings, revised prescribing guidelines or, in extreme cases, removal from the market.
So while detection works to a point, wouldn’t it be better if we could predict adverse drug events before a drug even hits the market?
Ben Reis and Aurel Cami of the Children’s Hospital Informatics Program (CHIP) thought so, and recently described a mathematical model for predicting drug-adverse event relationships that are likely to appear within a few years of a drug’s entry into the market, such as an increased risk of heart attack.
“This approach allows us to make the important transition from detection to prediction,” says Reis, who leads CHIP’s Predictive Medicine Group. “We can potentially identify a dangerous drug side effect early on, instead of having to wait for sufficiently many patients to be affected by it.”
The model builds networks of drug/adverse drug event (ADE) relationships and assigns probabilities for their actual likelihood. To do so, it combines available drug safety information, drug chemistry data (the intrinsic properties of the drug molecules) and data on drug and event taxonomy (to make sure the model understands that “heart attacks” in one data set are the same as “myocardial infarctions” in another).
“Given the myriad of drugs and complex relationships that exist in the pharmacological domain, we felt that a network-based approach would be a promising way to try to predict unforeseen but likely adverse events,” explains Cami. “For the approach to be properly validated, though, we knew we needed to use historical and current data on drug ADEs.”
For historical data to test the model, Cami and Reis used a snapshot of drug-ADE relationships in a commercially available database from Lexicomp for 2005. They ran that information through their model – only using chemical and taxonomic data available that same year – to generate a list of predicted side effects.
When they compared that list with a 2010 snapshot of the Lexicomp database, they found the model to be quite effective. Using just the data available in 2005, the model picked up more than 40 percent of the drug-side effect relationships reported between 2006 and 2010. It also correctly caught 95 percent of the relationships in the 2005 data that were later recognized as false-positives.
Cami and Reis both recognize that their model is just a first step. “Ideally, we’d like to see the creation of multiple models, each taking a different but complementary approach,” says Cami. “That way we could build consensus predictions about what events would be likely to occur, much like meteorologists combine multiple weather models to predict where a hurricane will likely make landfall.”
If we get to that point, the benefits will be enormous. “Today we rely mainly on post-marketing surveillance to identify unknown drug ADEs, especially with novel drug classes,” says Shannon Manzi, a pharmacist in Children’s Emergency Department and one of the study’s co-authors. “Being able to predict a potential relationship that wasn’t previously considered will improve the safety of drugs as they come to market, benefiting both drug companies and patients.”