Accurate and interpretable drug-drug interaction prediction enabled by knowledge subgraph learning.

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence

Background:

• Predicting drug-drug interactions (DDIs) is crucial but challenging due to the rarity of known interactions.
• Existing deep learning methods for DDI prediction often require extensive data, which is a limitation given the scarcity of known DDIs.

Purpose of the Study:

• To develop a novel graph neural network-based method, KnowDDI, for accurate and interpretable drug-drug interaction prediction.
• To address the challenge of limited known DDIs by enhancing drug representations and utilizing biomedical knowledge graphs.

Main Methods:

• KnowDDI employs a graph neural network architecture to enhance drug representations by integrating information from large biomedical knowledge graphs.
• It constructs a knowledge subgraph for each drug pair to interpret predicted DDIs, with edge strengths indicating interaction importance or similarity.
• The method implicitly compensates for the lack of known DDIs through enriched drug representations and propagated drug similarities.

Main Results:

• KnowDDI achieved state-of-the-art prediction performance on two benchmark DDI datasets.
• The method demonstrated improved interpretability in predicting DDIs.
• KnowDDI showed greater resilience to sparser knowledge graphs compared to existing methods, highlighting the importance of propagated drug similarities.

Conclusions:

• KnowDDI effectively integrates deep learning efficiency with the rich knowledge from biomedical graphs for DDI prediction.
• As an open-source tool, KnowDDI can be applied to various interaction prediction tasks, advancing biomedicine and healthcare.
• The approach demonstrates the utility of leveraging external knowledge to overcome data limitations in predictive modeling.