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

Biomedical knowledge relevant to disease mechanisms and therapeutic discovery is distributed across structured knowledge graphs (KGs) and unstructured text. Although pretrained language models provide strong semantic embeddings, adapting them to biomedical link prediction typically requires domain-specific fine-tuning and may weaken explicit structural constraints. An open challenge is to enhance link prediction with principled knowledge-language fusion while keeping language encoders lightweight and controllable.

We propose KLaR, a knowledge-language representation framework for biomedical KG link prediction. Given a query entity, KLaR encodes its local k-hop neighborhood with a relational GNN and constructs mechanism-consistent textual contexts by template-based textualization of random-walk paths within the same subgraph. These contexts are embedded using a frozen sentence-embedding model (without external retrieval or task-specific fine-tuning), and structural and textual views are aligned and fused via gated integration. To model heterogeneous biomedical interactions, KLaR uses a sparse mixture-of-experts decoder for triple scoring. On PharmKG, HetioNet, and DTINet, KLaR yields consistent gains over structure-only and KG-language hybrid baselines under standard filtered evaluation. We additionally report text-only LLM reference results under a candidate-restricted protocol due to the infeasibility of full-entity ranking for prompted generation. Case studies suggest that KLaR can recover biologically plausible disease-gene and drug-target associations missing from the original graphs, supporting hypothesis generation without domain-specific language model fine-tuning.

Code and processed data are available at: https://github.com/stardj/KLaR.