TK-DDI: Accurate and efficient drug-drug interaction prediction via token encoding

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Accurate drug-drug interaction (DDI) prediction is crucial for patient safety and preventing adverse drug events.
• Existing computational methods face challenges in modeling long-range molecular dependencies and identifying critical interaction substructures.

Purpose of the Study:

• To develop a novel deep learning framework, TK-DDI, for enhanced DDI prediction.
• To effectively model intramolecular dependencies and pinpoint salient substructures involved in drug interactions.

Main Methods:

• TK-DDI utilizes molecular tokenization to create unified drug representations, incorporating 2D and 3D information.
• A Transformer encoder learns contextual relationships between token pairs, capturing distant functional group influences.
• A two-stage attention strategy (intra-drug and inter-drug) elucidates interaction mechanisms by highlighting key substructures and fusing representations.

Main Results:

• TK-DDI demonstrates robust performance, outperforming existing state-of-the-art methods on benchmark datasets.
• The framework effectively models long-range intramolecular dependencies and identifies crucial substructures for DDI prediction.
• TK-DDI establishes a new benchmark for computational DDI prediction accuracy.

Conclusions:

• TK-DDI offers a significant advancement in DDI prediction accuracy and mechanistic understanding.
• The molecular tokenization and attention-based approach provide a powerful tool for drug safety assessment.
• This framework has the potential to enhance patient safety by enabling more reliable prediction of adverse drug events.