MSDAFL: molecular substructure-based dual attention feature learning framework for predicting drug-drug interactions

Area of Science:

• Pharmacology
• Computational Chemistry
• Bioinformatics

Background:

• Drug-drug interactions (DDIs) pose risks to patient safety and treatment efficacy.
• Predicting DDIs computationally is crucial for proactive risk assessment.
• Existing deep learning methods often neglect crucial substructure interaction information.

Purpose of the Study:

• To develop an advanced deep learning framework for enhanced DDI prediction.
• To leverage substructure interactions within drug pairs for improved model performance.
• To address limitations in current DDI prediction methodologies.

Main Methods:

• Introduced the molecular Substructure-based Dual Attention Feature Learning (MSDAFL) framework.
• Utilized self-attention and interactive attention modules to capture substructure information and interactions.
• Employed cosine similarity for interaction feature extraction and normalization to prevent overfitting.

Main Results:

• MSDAFL achieved high precision scores (e.g., 0.9707, 0.9991, 0.9987) and AUC scores (e.g., 0.9874, 0.9934, 0.9974) across multiple datasets.
• Cross-validation and cross-datum studies confirmed the model's robust performance in DDI prediction.
• The framework effectively utilizes inter-drug substructure information for superior prediction.

Conclusions:

• MSDAFL demonstrates significant potential for accurate and reliable DDI prediction.
• The substructure-based attention mechanism enhances the understanding of drug pair interactions.
• This approach contributes to improving drug safety and efficacy through advanced computational prediction.