Modeling polypharmacy side effects with graph convolutional networks.

Area of Science:

• Computational biology
• Pharmacogenomics
• Artificial intelligence in medicine

Background:

• Polypharmacy is common for complex diseases but increases the risk of adverse drug events due to drug-drug interactions.
• Predicting these interactions is challenging due to their rarity and limited observation in clinical trials, impacting patient morbidity and mortality.

Purpose of the Study:

• To present Decagon, a novel approach for modeling and predicting polypharmacy side effects.
• To accurately identify specific side effects associated with drug combinations.

Main Methods:

• Constructed a multimodal graph integrating protein-protein interactions, drug-protein targets, and drug-drug interactions representing side effects.
• Developed a graph convolutional neural network tailored for multirelational link prediction in multimodal networks.

Main Results:

• Decagon accurately predicts polypharmacy side effects, outperforming existing methods by up to 69%.
• The model learns representations of side effects and effectively models those with a molecular basis.
• Achieved good performance on non-molecular side effects through parameter sharing.

Conclusions:

• Decagon offers a powerful tool for predicting polypharmacy side effects, improving patient safety.
• Enables leveraging large-scale pharmacogenomic and patient data to identify and prioritize potential adverse drug events for further study.