Researchers at Stanford University built Decagon, a model that treated drugs, proteins and known side effects as one giant network, then tried to predict which side effects would appear in people taking two drugs together, a combination too rare and too dangerous to test directly in a trial.
The network combined protein-to-protein interactions, drug-to-protein targets and 964 types of documented drug-pair side effects, then learned the patterns that connect them. Tested against side effects already known to doctors but hidden from the model during training, Decagon correctly identified them up to 69 percent more often than the next-best computational method.
Decagon never ran a single clinical trial. It worked entirely from existing records, no new patients, no new cells, no new animals, just a better map of how drugs, proteins and known reactions connect. The predictions still need confirming case by case before a doctor acts on them, but the approach gave researchers a ranked list of which untested drug pairs are most worth checking first.
Read the original study
Zitnik, Agrawal and Leskovec, Bioinformatics, 2018 · doi.org/10.1093/bioinformatics/bty294
A map of risk nobody had to test.
Source
- Zitnik, M., Agrawal, M., & Leskovec, J. (2018). Modeling polypharmacy side effects with graph convolutional networks. Bioinformatics, 34(13), i457–i466. https://doi.org/10.1093/bioinformatics/bty294