Automating candidate gene prioritization with large language models: from naive scoring to literature-grounded...

Area of Science:

• Biomedical research
• Computational biology
• Genomics

Background:

• Transcriptomic studies generate vast gene data, posing challenges for identifying therapeutic targets.
• Large language models (LLMs) show promise for gene prioritization but often hallucinate and lack validation.
• Systematic validation against expert knowledge is crucial for reliable gene prioritization.

Purpose of the Study:

• To develop and validate a computational framework for systematic gene prioritization using LLMs and literature.
• To identify high-confidence therapeutic gene targets for sepsis.
• To overcome the limitations of LLMs in biomedical research through rigorous validation.

Main Methods:

• A two-stage framework combining LLM-based screening with literature validation was developed.
• Genes were screened for sepsis relevance using multi-criteria evaluation and retrieval-augmented generation from sepsis publications.
• A faithfulness evaluation system ensured LLM predictions aligned with literature evidence.

Main Results:

• The framework identified 609 sepsis-relevant genes with high filtering efficiency, enriched in inflammatory pathways.
• 30 ultra-high confidence therapeutic candidates were identified, including novel targets and known sepsis genes.
• Benchmark validation achieved 71.2% recall, correlating computational confidence with evidence quality.

Conclusions:

• The framework transforms unreliable LLM outputs into systematically validated biological insights.
• It provides a practical tool for prioritizing experimental validation and literature-guided biomarker discovery.
• The modular design allows adaptation to other diseases, offering a versatile approach.