Could AI Predictions Make Polypharmacy Pharmacovigilance Prophylactic?
What if pharmacovigilance could identify dangerous drug interactions before patients ever experience them? Recent breakthroughs suggest AI could predict and potentially prevent adverse reactions from drug combinations.
DCSE (Drug Combinations Side Effects) is a machine learning model developed by researchers Ruben Jimenez and Alberto Paccanaro to predict adverse reactions from combinations of different medicines. They seek to address a longstanding challenge in drug safety: clinical trials realistically cannot investigate every possible drug combination, so many potentially harmful interactions only emerge when these medicines are combined in larger populations.
“DCSE-twin represents an opportunity to use patient population data to flag and prioritise adverse effects caused by drug combinations for further validation”
Dr Ruben Jimenez, researcher and co-author of the 2026 DCSE study
DCSE works by learning numerical representations, called latent signatures, of individual medicines, drug combinations and adverse events. By identifying known patterns within existing drug safety data, the model can estimate the likelihood of specific side effects occurring when two medicines are taken together. DCSE can even make predictions for drug combinations with no previously documented adverse reactions.
To test its predictive capabilities, researchers trained the model on historical data and assessed whether it could predict side effects reported between 2009 and 2014. Impressively, DCSE consistently outperformed competing methods, even when evaluating previously uncharacterised combinations.
Although traditional pharmacovigilance has focused on identifying safety signals from adverse reactions that have already happened, predictive models like DCSE could invite an additional preventative layer, allowing safety teams to investigate potentially dangerous combinations before patient exposure.
These predictions could inform risk mitigation strategies, guide the design of clinical trials and prioritise combinations requiring further investigation. Instead of being faced with the insurmountable task of investigating every possible combination, clinical trials could just investigate combinations which algorithms predict could be problematic. Perhaps in the future, validated systems could also support prescribing decisions and warn clinicians about potential interactions before treatment begins.
Of course, these predictions are not gospel. A model’s identification of an association cannot establish causality, and reliable clinical deployment requires much more validation and careful consideration of false alarms. However, perhaps the most interesting insight from this innovation is the possibility of pharmacovigilance evolving from a mainly reactive discipline into an anticipatory one. The ultimate ambition is not simply to detect adverse reactions faster, but to stop them from happening at all.
