Predicting Drug-Drug Interaction via Dual-Drug Visual Representation

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Drug-drug interaction (DDI) prediction is crucial for medication safety.
• Existing DDI models often neglect visual spatial and structural information from molecules.
• There is a need for advanced methods that capture the visual interface of drug pairs.

Purpose of the Study:

• To propose DDVR-DDI, a novel vision-based framework for predicting DDIs.
• To leverage visual molecular representations for enhanced DDI prediction accuracy.
• To improve the interpretability of DDI prediction models.

Main Methods:

• Encoding drug pairs as a single fused molecular image.
• Utilizing a two-stage self-supervised pretraining strategy (position-invariant contrastive learning and jigsaw puzzle task).
• Implementing a multiexpert voting mechanism for ensemble inference.

Main Results:

• Achieved state-of-the-art performance on benchmark DDI datasets.
• Demonstrated enhanced prediction accuracy and stability through ensemble methods.
• Validated model interpretability using Grad-CAM visualizations and case studies.

Conclusions:

• Vision-based modeling offers a promising approach for accurate DDI prediction.
• The DDVR-DDI framework provides mechanistic insights into drug interactions.
• The model successfully identified chemically significant substructures in drug interactions.