Predicting Drug-Drug Interactions Using Heterogeneous Graph Neural Networks: HGNN-DDI

This study presents a novel approach to predicting drug-drug interactions (DDIs) using heterogeneous graph neural networks (HGNNs) augmented with attention mechanisms. In the proposed framework, drugs and proteins are represented as nodes within a heterogeneous graph, where edges encode drug-drug, drug-protein, and protein-protein interactions.

To capture chemical structure information, we leverage ChemBERTa, a pretrained language model that processes SMILES representations of drugs. Similarities between drug structures are quantified using RDKit. The HGNN model integrates multiple biological modalities—chemical structures, drug-protein interactions (DPIs), and protein-protein interactions (PPIs)—to learn richer node embeddings.

A Multi-Layer Perceptron (MLP) classifier is employed to predict the type of interaction between drug pairs, producing probability scores for each potential DDI. The model improves prediction accuracy by incorporating complementary biological data and offers interpretable outputs that are valuable for pharmacovigilance and personalized treatment planning.

These predictions can aid healthcare professionals in evaluating the safety and efficacy of drug combinations, especially for patients with complex or evolving medical conditions.

Recommended citation: Liu, H., Li, S., & Yu, Z. (2024). Predicting drug-drug interactions using heterogeneous graph neural networks: HGNN-DDI. Applied and Computational Engineering, 79, 77–89.
Download Paper