Clinical-Grade Somatic Variant Interpretation Performance via a Rule-Constrained Large Language Model Framework...

Area of Science:

• Genomics
• Computational Biology
• Clinical Pathology

Background:

• Somatic variant interpretation in oncology is complex, requiring integration of gene biology, tumor context, and diverse evidence.
• Large language models (LLMs) offer potential for variant interpretation but raise concerns about reproducibility and safety.

Purpose of the Study:

• To develop and evaluate OLIVE (Oncology Logic-Informed Variant Evaluator), a rule-constrained LLM-based decision support framework for clinical somatic variant interpretation.
• To assess OLIVE's performance and concordance with historical laboratory classifications in a real-world molecular pathology setting.

Main Methods:

• OLIVE was developed to summarize data sources and apply explicit gene-specific guidance files for structured variant classification.
• Performance was evaluated on 200 clinical tumor NGS cases (1,437 variant observations) comparing OLIVE's classifications with historical laboratory interpretations.
• Concordance was assessed for Pathogenic/Likely Pathogenic versus Variant of Uncertain Significance (VUS) determinations across three replicates.

Main Results:

• Mean concordance with historical interpretations was high at 97.5% across three replicates.
• Only 2.9% of variants showed discordance, often reflecting inherent interpretive ambiguity or evolving evidence rather than model instability.
• Expert adjudication of discordant variants showed equal preference for laboratory or OLIVE classifications, indicating OLIVE supports expert decision-making.

Conclusions:

• The OLIVE framework demonstrates reproducible support for expert somatic variant interpretation in clinical molecular pathology.
• Its performance is primarily driven by expert-defined guidance, suggesting robustness and adaptability across different LLMs.
• OLIVE enhances the reliability and efficiency of somatic variant interpretation in oncology.