MEDL-DDI: Example-Driven Learning With Multi-Source Features for Predicting Drug-Drug Interaction.

Area of Science:

• Pharmacology
• Computational Chemistry
• Bioinformatics

Background:

• Accurate drug-drug interaction (DDI) prediction is essential for safe and effective combination therapies.
• Current DDI prediction methods often fail to integrate structural and sequential drug information, leading to suboptimal performance.
• Class imbalance in datasets poses a significant challenge, biasing prediction models against less frequent interactions.

Purpose of the Study:

• To develop a robust multimodal fusion framework for DDI prediction that integrates structural and sequential drug information.
• To address the challenge of class imbalance in DDI prediction datasets.
• To enhance the accuracy and reliability of drug-drug interaction predictions.

Main Methods:

• Proposed a multi-source example-driven learning framework for DDI (MEDL-DDI).
• Integrated structural and sequential drug representations through multimodal fusion.
• Utilized chemical knowledge enrichment, Transformer for global semantic features, graph information bottleneck for substructures, and an example-driven mechanism to handle class imbalance.

Main Results:

• MEDL-DDI demonstrated superior performance compared to state-of-the-art methods on three benchmark datasets.
• The framework effectively integrated multimodal drug information for enhanced prediction accuracy.
• The example-driven mechanism successfully mitigated prediction bias caused by class imbalance.

Conclusions:

• MEDL-DDI offers a powerful approach for accurate drug-drug interaction prediction by leveraging multimodal data fusion and addressing class imbalance.
• The framework shows significant potential for optimizing combination therapies and improving patient safety.
• A case study on cardiovascular drugs confirmed the practical value and applicability of MEDL-DDI.