KLaR: fusing knowledge graphs and language models for biomedical target discovery.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Artificial Intelligence

Background:

• Biomedical knowledge is fragmented across structured knowledge graphs (KGs) and unstructured text.
• Existing language models struggle to integrate KG structure with text embeddings for link prediction.
• A key challenge is fusing knowledge and language representations efficiently without compromising KG structure.

Purpose of the Study:

• To develop a novel knowledge-language representation framework (KLaR) for biomedical KG link prediction.
• To enhance link prediction by integrating structural KG information with textual context.
• To enable effective knowledge-language fusion while maintaining lightweight and controllable language encoders.

Main Methods:

• KLaR encodes local KG neighborhoods using a relational Graph Neural Network (GNN).
• Mechanism-consistent textual contexts are generated via template-based textualization of random walks.
• Structural and textual views are fused using gated integration with a frozen sentence-embedding model.
• A sparse mixture-of-experts decoder is employed for heterogeneous biomedical interaction scoring.

Main Results:

• KLaR demonstrated consistent performance gains on PharmKG, HetioNet, and DTINet compared to baseline methods.
• The framework successfully recovered biologically plausible missing associations, aiding hypothesis generation.
• Case studies showed KLaR's ability to identify novel disease-gene and drug-target relationships.

Conclusions:

• KLaR offers an effective approach for knowledge-language fusion in biomedical link prediction.
• The framework supports hypothesis generation for drug discovery without requiring domain-specific language model fine-tuning.
• KLaR advances the integration of structured and unstructured biomedical data for enhanced discovery.