Multisource Attention-Mechanism-Based Encoder-Decoder Model for Predicting Drug-Drug Interaction Events.

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Drug-drug interactions (DDIs) pose significant clinical challenges.
• Existing computational methods often rely on limited, single-source drug features, hindering accurate prediction.
• Effective drug representation is crucial for understanding and predicting complex interactions.

Purpose of the Study:

• To propose novel attention-mechanism-based encoder-decoder models for predicting DDIs and DDI-associated events.
• To enhance drug representation by incorporating multisource information.
• To improve the accuracy and generalization of DDI prediction models.

Main Methods:

• Developed two models: MAEDDI for DDI prediction and MAEDDIE for DDI-associated event prediction.
• Integrated multisource drug features using sequence-based, structural, biochemical, and statistical data.
• Employed self-attention, cross-attention, and graph attention networks for feature fusion.
• Constructed a multisource feature fusion network for comprehensive drug representation.

Main Results:

• MAEDDI and MAEDDIE outperformed state-of-the-art methods in DDI and DDI-associated event prediction tasks.
• Visualization analysis confirmed effective semantic feature learning and good drug representation.
• MAEDDIE successfully identified 43 DDI events for favipiravir with high accuracy.
• The models demonstrated strong generalization capabilities in predictions.

Conclusions:

• Multisource feature integration and attention mechanisms significantly improve DDI prediction accuracy.
• The proposed models provide a robust framework for characterizing complex drug-drug relationships.
• These computational tools can aid in screening potential DDI events and enhancing drug safety.