从真实世界MTB数据中提取基于大语言模型的分子层面患者相似性

Background: Molecular tumor boards (MTBs) frequently evaluate patients with rare molecular profiles where prospective evidence is scarce. To address this challenge, we investigated whether large language models (LLMs) can structure real-world MTB documentation into analyzable representations to identify molecularly similar patients. Methods: MTB documentation from 2,788 patients and pathology reports from 506 patients discussed at the Charité MTB (2020-2025) were structured using NVIDIA Nemotron-49B. Textual summaries and structured features on prior targeted therapies, immunohistochemistry (IHC), and molecular alterations (MOL; CiViC-weighted) were extracted, yielding 1,736 textual, binary and numeric features. The LLM output was vector embedded using bge-multilingual-gemma2. Patient similarity was computed using cosine distance for embeddings, and Jaccard (binary) and absolute error (numeric) for structured features. Similarity rankings were ensembled via reciprocal-rank fusion and evaluated using MRR and nDCG, against BM25 and embedding baselines on raw MTB documentation. As a proxy for ground truth, MTB recommendations were structured into course of action, drug, agent class, and clinical trial elements and used as similarity targets. Results: LLM-based information extraction achieved F1 scores of 96% for mutation detection and 92% for IHC, evaluated against 30 manually annotated pathology reports. The ensemble similarity method - combining structured-feature similarity, summary-embedding similarity, and BM25 via reciprocal-rank fusion - showed the highest alignment with MTB recommendations (MRR@1000 = 25.8, nDCG@10 = 10.1), outperforming BM25 alone (22.8, 8.6) or text-embedding similarity of MTB documentation (22.2, 8.3), with all improvements statistically significant (p < 0.01). Structured features derived from MTB documentation (20.2, 9.2) outperformed those derived from pathology reports (17.9, 8.0). Together, these findings indicate that LLM-derived representation improved therapy-aligned patient similarity over text- and embedding-based baselines. Across all cases, 486 unique therapeutic entities were identified; among 290 drug entities, 163 were recommended to more than one patient, and 65% of patients shared at least one recommendation with another case. Conclusion: LLM-derived clinical-molecular representations enabled scalable retrieval of molecularly matched cases in real-world MTB datasets. This approach supports institutional case library formation and systematic case series generation, enhancing evidence generation in precision oncology. [O.S. and S.L. contributed equally to this work.]