Clinical-Grade Somatic Variant Interpretation Performance via a Rule-Constrained Large Language Model Framework (Oncology Logic-Informed Variant Evaluator).

Interpretation of somatic variants in clinical oncology requires integration of gene-specific biology, tumor context, and multiple evidence sources. Although large language models (LLMs) can assist variant interpretation, concerns remain regarding reproducibility, safety, and dependence on opaque model behavior. A rule-constrained LLM-based decision support framework, Oncology Logic-Informed Variant Evaluator (OLIVE), was developed and evaluated for somatic variant interpretation in a clinical molecular pathology service. OLIVE summarizes multiple data sources and applies explicit gene-specific guidance files to generate structured prompts for final classification and interpretation. Performance was assessed on 200 consecutive clinical tumor next-generation sequencing cases comprising 1437 variant observations (1228 unique variants). Concordance with historical laboratory classifications was evaluated for reportable (pathogenic/likely pathogenic) versus variant of uncertain significance determinations. Across three replicates, mean concordance with historical interpretation was 97.5% (range, 97.4% to 97.6%). Forty-two variants (2.9%) showed discordance in at least one replicate, with most discordances being consistent across runs. Blinded post hoc expert adjudication of discordant variants demonstrated interpretive variability favoring the laboratory or OLIVE classification in equal proportions (21 versus 21 variants). Discordances predominantly reflected intrinsic interpretive ambiguity, evidence evolution, and rule stringency rather than model instability. Similar results with a different LLM indicate performance is not model dependent and is primarily driven by expert guidance. These results indicate that OLIVE can reproducibly support expert somatic variant interpretation in a real-world clinical setting.