MEGH-DDI: MolT5-driven edge-aware graph learning with Hop2Token representation for drug-drug interaction prediction.

MEGH-DDI achieves 98.36% accuracy in drug-drug interaction prediction on the DrugBank dataset, surpassing previous methods and achieving over 99.7% AUC and AP.

• Why it matters: Accurate DDI prediction is vital for drug safety and development, but existing molecular representations often miss dynamic substructure and semantic interactions, limiting predictive performance.
• What they did: The study introduces MEGH-DDI, a multimodal framework combining Mol2vec and MolT5 embeddings, dynamic feature selection, edge-aware GCNII with bond-level attention, and Hop2Token for high-order substructure modeling, tested on two datasets with both transductive and inductive settings.
• The result: MEGH-DDI demonstrates high accuracy and F1 scores (99.05% on ChChMiner, outperforming previous methods by 2.74 points), and shows potential for generalizing to unseen drugs and complex DDI scenarios, with interpretability through atom, bond, and path analysis.