TRACE-DDI: A Hybrid Framework of Transformer-GAT Context Encoder and Pathway-Anchored Knowledge Graphs for DDI...

Area of Science:

• Pharmacology
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Accurate drug-drug interaction (DDI) prediction is vital for patient safety and efficient drug development.
• Existing models often fail to integrate chemical structures and biological knowledge effectively, limiting their scope.
• A multiscale biological context is essential for comprehensive DDI prediction.

Purpose of the Study:

• To introduce TRACE-DDI, a hybrid framework for DDI prediction.
• To integrate diverse biological data modalities for improved prediction accuracy.
• To provide biologically interpretable insights for predicted DDIs.

Main Methods:

• Developed TRACE-DDI, a unified framework integrating SMILES sequences, molecular graph topology, and pathway-anchored knowledge graphs.
• Employed a Transformer-Graph Attention Network (GAT) cross-modal context encoder for feature propagation across heterogeneous data.
• Utilized multilayer graph-based feature propagation for fine-grained chemical understanding and pathway-level context aggregation.

Main Results:

• TRACE-DDI significantly outperformed state-of-the-art baseline models on a benchmark dataset across key evaluation metrics.
• The framework demonstrated consistent improvements in predictive accuracy.
• Pathway-level subgraph visualization and centrality analysis identified key biological nodes relevant to predicted DDIs.

Conclusions:

• TRACE-DDI offers a unified and interpretable approach to biologically grounded DDI prediction.
• The framework successfully integrates multiscale biological context for enhanced DDI prediction.
• TRACE-DDI provides hypothesis-generating biological insights, advancing AI-driven pharmacology.