A multimodal deep learning framework for predicting drug-drug interaction events.

Area of Science:

• Pharmacology
• Computational Biology
• Artificial Intelligence

Background:

• Drug-drug interactions (DDIs) pose significant risks in pharmaceutical research.
• Current DDI prediction methods often focus on interaction presence rather than specific associated events.
• Understanding DDI events is key to elucidating mechanisms and adverse reactions.

Purpose of the Study:

• To develop a novel framework for predicting specific DDI-associated events.
• To leverage multimodal drug features for enhanced DDI event prediction accuracy.
• To provide a more informative approach beyond simple DDI detection.

Main Methods:

• Collected DDIs from the DrugBank database.
• Extracted 65 categories of DDI events using dependency analysis and event trimming.
• Proposed DDIMDL, a multimodal deep learning framework utilizing chemical substructures, targets, enzymes, and pathways.
• Employed deep neural networks (DNNs) for sub-models and a joint DNN for cross-modality representation learning.

Main Results:

• DDIMDL achieved high accuracy and efficiency in predicting DDI events.
• The model outperformed state-of-the-art and baseline DDI event prediction methods.
• Chemical substructures were identified as the most informative drug features.
• Combining substructures, targets, and enzymes yielded an accuracy of 0.8852 and an AUC-PR of 0.9208.

Conclusions:

• DDIMDL offers a powerful and accurate approach for DDI event prediction.
• The multimodal deep learning strategy effectively integrates diverse drug features.
• This framework advances the understanding of DDI mechanisms and potential adverse reactions.