Causal knowledge graph analysis identifies adverse drug effects.

Area of Science:

• Biomedical Informatics
• Causal Inference
• Knowledge Representation

Background:

• Biomedical knowledge graphs and causal models are valuable but disconnected.
• Knowledge graphs lack probabilistic semantics; causal models lack background knowledge integration.

Purpose of the Study:

• To bridge the gap between knowledge graphs and causal models.
• To develop Causal Knowledge Graphs (CKGs) for principled causal inference and hypothesis formulation.

Main Methods:

• Extended knowledge graphs with formal causal semantics, creating Causal Knowledge Graphs (CKGs).
• Constructed a Drug-Disease CKG (DD-CKG) integrating disease pathways, drug indications, and side effects.
• Applied DD-CKG to UK Biobank and MIMIC-IV for large-scale mediation analysis of drug effects.

Main Results:

• CKGs enabled deconfounding and hypothesis formulation using background knowledge.
• Successfully reproduced known adverse drug reactions and identified novel candidate effects.
• Improved prediction of shared drug indications by combining predicted effects with existing databases.

Conclusions:

• CKGs provide a generalizable, knowledge-driven framework for scalable causal inference.
• The methodology demonstrates clinical relevance and supports automated large-scale mediation analysis.
• This approach enhances understanding of drug effects and disease progression.