AttentionDDI: Siamese attention-based deep learning method for drug-drug interaction predictions

Area of Science:

• Pharmacology and Bioinformatics
• Artificial Intelligence in Medicine

Background:

• Drug-drug interactions (DDIs) can cause unexpected side effects.
• Experimental testing of all drug combinations is infeasible.
• Machine learning offers a scalable approach to predict DDIs.

Purpose of the Study:

• To develop an advanced machine learning model for predicting drug-drug interactions.
• To integrate diverse drug similarity measures for improved DDI prediction accuracy.

Main Methods:

• A Siamese self-attention multi-modal neural network was proposed.
• The model integrates drug similarity derived from targets, pathways, and gene expression.
• End-to-end training was employed for model optimization.

Main Results:

• The model achieved comparable or superior prediction performance (AUPR 0.77-0.92) on benchmark datasets.
• An attention mechanism provided model explainability by highlighting salient features.
• Novel DDI predictions were validated using independent data resources.

Conclusions:

• Siamese multi-modal neural networks can accurately predict DDIs.
• Attention mechanisms enhance the explainability of DDI prediction models.
• This approach offers a powerful tool for drug safety assessment.