Modeling polypharmacy side effects with graph convolutional networks

• DOI: 10.1093/bioinformatics/bty294
• Publication Date: 2018-04-12T19:32:51Z
• Authors: Marinka Zitnik; Monica Agrawal; Jure Leskovec
• Keywords: FOS: Computer and information sciences; 0301 basic medicine; Computer Science - Machine Learning; Drug-Related Side Effects and Adverse Reactions; Molecular Networks (q-bio.MN); Machine Learning (stat.ML); Models, Biological; Machine Learning (cs.LG); 03 medical and health sciences; Statistics - Machine Learning; Humans; Quantitative Biology - Molecular Networks; Drug Interactions; Protein Interaction Maps; Ismb 2018–Intelligent Systems for Molecular Biology Proceedings; 0303 health sciences; Data Visualization; Computational Biology; 3. Good health; FOS: Biological sciences; Polypharmacy; Female; Neural Networks, Computer; Software

The use of drug combinations, termed polypharmacy, is common to treat patients with complex diseases or co-existing conditions. However, a major consequence polypharmacy much higher risk adverse side effects for the patient. Polypharmacy emerge because drug-drug interactions, in which activity one may change, favorably unfavorably, if taken another drug. knowledge interactions often limited these relationships are rare, and usually not observed relatively small clinical testing. Discovering thus remains an important challenge significant implications patient mortality morbidity.Here, we present Decagon, approach modeling effects. constructs multimodal graph protein-protein drug-protein target effects, represented as where each effect edge different type. Decagon developed specifically handle such graphs large number types. Our develops new convolutional neural network multirelational link prediction networks. Unlike approaches predicting simple interaction values, can predict exact effect, any, through given combination manifests clinically. accurately predicts outperforming baselines by up 69%. We find that it automatically learns representations indicative co-occurrence patients. Furthermore, models particularly well have strong molecular basis, while on predominantly non-molecular achieves good performance effective sharing model parameters across opens opportunities pharmacogenomic population data flag prioritize follow-up analysis via formal pharmacological studies.Source code preprocessed datasets at: http://snap.stanford.edu/decagon.