Large Language Models Meet Biomedical Knowledge Graphs for Mechanistically Grounded Therapeutic Prioritization

Area of Science:

• Computational Biology
• Pharmacology
• Bioinformatics

Background:

• Current drug repurposing methods struggle to differentiate biologically plausible candidates from those with historical associations.
• Existing approaches lack robust mechanisms for prioritizing novel therapeutic applications based on underlying biology.

Purpose of the Study:

• To introduce DrugKLM, a hybrid framework integrating biomedical knowledge graphs and large language model (LLM) mechanistic reasoning.
• To enable mechanistically grounded therapeutic prioritization for drug repurposing.
• To improve the identification of biologically plausible drug candidates.

Main Methods:

• Developed DrugKLM, a framework combining knowledge graph structure with LLM-based mechanistic reasoning.
• Evaluated DrugKLM on benchmark datasets against knowledge graph-only and LLM-only baselines.
• Assessed the functional alignment of DrugKLM confidence scores with molecular phenotypes and survival data.
• Conducted expert curation across five cancer types to analyze prioritization behavior.

Main Results:

• DrugKLM outperformed existing baselines, including TxGNN, in therapeutic prioritization.
• DrugKLM confidence scores showed functional alignment with molecular phenotypes, correlating with improved survival in 12 TCGA cancers.
• The framework preferentially captured biologically perturbational signals over historical indication patterns.
• Expert curation confirmed DrugKLM's ability to prioritize candidates with coherent mechanistic rationale and clinical context.

Conclusions:

• DrugKLM is an effective evidence-integrative framework for therapeutic prioritization in drug repurposing.
• The framework translates heterogeneous biomedical data into mechanistically interpretable and clinically relevant hypotheses.
• DrugKLM advances the field by providing mechanistically grounded guidance for identifying novel drug candidates.