Enabling Drug-Drug Interaction Event Prediction with Multi-view-enhanced Chemical Structural Information

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in medicine

Background:

• Drug-drug interactions (DDIs) pose significant risks to patient safety and drug efficacy.
• Current deep learning models for DDI prediction often overlook crucial chemical structure details and diverse interaction types.
• This limitation hinders accurate prediction and optimization of therapeutic outcomes.

Purpose of the Study:

• To develop an advanced deep learning framework, ChemDDI, for more accurate DDI event prediction.
• To effectively integrate multi-view chemical structural information and multi-relational interaction data.
• To enhance the prediction of both common and rare drug-drug interaction events.

Main Methods:

• ChemDDI utilizes multi-view graph- and image-based encoders to extract rich chemical structural features from 3D drug conformations.
• Transformer-based graph neural networks and relational graph embeddings are employed to model multi-relational interaction information.
• Contrastive learning is incorporated to align interaction features, improving the robustness of DDI prediction.

Main Results:

• ChemDDI significantly outperforms existing state-of-the-art methods in DDI event prediction.
• The framework demonstrates substantial improvements in predicting rare drug-drug interaction events.
• Experimental validation confirms the efficacy of the multi-view structural information integration.

Conclusions:

• ChemDDI offers a robust and accurate deep learning approach for DDI prediction by leveraging comprehensive chemical structure and interaction data.
• The framework holds potential for improving patient safety and accelerating drug discovery processes.
• The developed model provides a valuable tool for understanding and predicting complex drug interactions.