Graph neural networks for drug-drug interaction prediction in polypharmacy patients

Abstract

We propose a heterogeneous graph neural network model for predicting adverse drug-drug interactions in patients taking multiple medications. The model encodes drug molecular structures, protein targets, and known side effects as a knowledge graph with 1.2 million edges. Evaluated on the DrugBank and TWOSIDES datasets, our approach achieves AUROC of 0.934 for predicting 963 distinct polypharmacy side effect types, outperforming existing methods by 4.7%.

Topics

Full Text Preview

The heterogeneous knowledge graph contains 645 drug nodes, 19,085 protein nodes, and 963 side-effect nodes connected by drug-protein, protein-protein, and drug-side-effect edges. Drug nodes are initialized with Morgan fingerprint features (2048-bit). A 3-layer R-GCN with relation-specific weight matrices learns 256-dimensional node embeddings. Drug-drug interaction prediction is formulated as a link prediction task using a DistMult decoder. Training uses binary cross-entropy loss with negative sampling ratio of 5:1.

Published Through