MUFFIN

MUFFIN predicts drug-drug interactions by integrating multi-scale features from drug molecular structures and large-scale biomedical knowledge graphs using a deep learning model.

Key Features:

• Multi-scale Feature Fusion: Employs a deep learning model to integrate multi-modal data including molecular structure information and semantic knowledge from biomedical networks for a comprehensive representation of drug interactions.
• Bi-level Cross Strategy: Implements a bi-level cross strategy with cross- and scalar-level components to fuse features derived from large-scale knowledge graphs (KGs) and drug molecular structures.
• Integration of Drug-Self Structure and KG Information: Combines drug molecular graphs with KG-derived biomedical relations involving genes, diseases, and pathways to capture intrinsic drug properties and inter-entity relationships.
• Handling Limited Labeled Data: Mitigates limited labeled data issues by leveraging complementary features from drug molecular graphs and extensive KGs to enhance model generalization.

Scientific Applications:

• Binary-class DDI prediction: Predicts presence or absence of DDIs and has been evaluated on binary-class datasets.
• Multi-class DDI prediction: Predicts categories of DDIs and has been evaluated on multi-class datasets.
• Multi-label DDI prediction: Predicts multiple concurrent interaction labels per drug pair and has been evaluated on multi-label datasets, reporting performance exceeding state-of-the-art baselines.

Methodology:

Uses a deep learning model that integrates multi-modal data via a bi-level cross strategy with cross- and scalar-level components to fuse features from drug molecular graphs and large-scale knowledge graphs.