PO.MD01.02 · 分子诊断与数据

用于肿瘤基因组解读的大语言模型

Large language models for tumor genomic interpretation

编号 4100 展板 5 时间 4/21 09:00–12:00 区域 Section 1 主讲 Jennifer Yu, MS

分会场 AACR Project GENIE: Genomic Characterization

查看 PDF 下载 PDF 🔒 查看 / 下载完整 PDF 需登录并开通下载套餐 · 查看套餐 / 开通 AACR 官方页面

作者与单位 Authors & Affiliations

Jennifer Yu¹, Madison Darmofal², Michele Waters², January Choy², Thinh N. Tran², Chenlian Fu¹, Leah Morales¹, Kaicheng U¹, Ross L. Levine², Nikolaus Schultz², Michael F. Berger², Quaid Morris², Justin Jee²

¹Tri-Institutional Computational Biology & Medicine, Weill Cornell Medicine, New York, NY,²Memorial Sloan Kettering Cancer Center, New York, NY

摘要 Abstract

中文摘要

引言：在真实世界数据上训练的算法有助于肿瘤基因组预测任务，例如识别癌症驱动突变和推断癌症类型。在大型自然语言语料库上训练的通用大语言模型（LLM）能在多大程度上通过零样本推理替代或补充此类领域特异性算法，目前尚不清楚。方法：我们评估了专有模型（GPT-5、o3-mini、GPT-4o 和 Claude 3.7 Sonnet）、开放权重模型（DeepSeek 和 Qwen3）以及领域专用医学模型（MedGemma）LLM 在三项任务上的零样本表现：（i）在具有配对肿瘤-全血谱分析的患者中区分肿瘤体细胞突变与克隆性造血（CH）变异（N=37,179 例患者；54,807 例样本），（ii）以 OncoKB 数据集作为阳性对照对致癌变异进行分类（N=10,489 例患者；10,752 例样本；13,470 个变异），以及（iii）在多机构 AACR GENIE 数据集中根据肿瘤基因组谱预测癌症类型（N=97,074 例患者；102,791 例样本）。结果：多个 LLM 接近了 MetaCH（一个用于区分体细胞肿瘤突变与 CH 变异的监督模型）的准确率。o3-mini 在区分致癌驱动突变与良性乘客突变方面取得了最高准确率。在具有 KEAP1 突变的非小细胞肺癌患者中，那些其 VUS 被 GPT-5 分类为致癌性的患者的总生存期比 VUS 被分类为良性的患者更差。GPT-5、o3-mini 和 Claude 3.7 Sonnet 在利用肿瘤基因组谱对 34 种癌症类型进行分类时，其准确率接近于监督模型 GDD-ENS。将 GPT-5 和 GDD-ENS 的预测结果相结合的集成方法使跨机构泛化能力和性能平均提高了 20%。在其推理过程中，LLM 讨论了与 GDD-ENS 特征重要性相一致的临床相关基因组特征。结论：在无需针对特定任务训练的情况下，LLM 在所有任务中均取得了与专用监督模型相当的性能。三项癌症基因组学预测任务的 F1 分数比较。突变状态（肿瘤体细胞 vs. CH）致癌变异（良性 vs. 致癌）癌症类型（34 种类型）GPT-5 0.94 0.80 0.49 GPT-4o 0.96 0.75 0.32 o3-mini 0.93 0.82 0.43 Claude 3.7 Sonnet 0.95 0.79 0.43 DeepSeek 0.86 0.79 0.23 Qwen3 0.95 0.79 0.26 MedGemma 0.86 0.77 0.19 MetaCH 0.98 n/a n/a AlphaMissense n/a 0.90 n/a GDD-ENS n/a n/a 0.57

查看英文原文 English abstract

Introduction: Algorithms trained on real-world data aid in tumor genomic prediction tasks, such as identifying cancer driver mutations and inferring cancer type. The extent to which generalist large language models (LLMs) trained on large natural language corpora can replace or supplement such domain-specific algorithms with zero-shot inference is unknown. Methods: We evaluated the zero-shot performance of proprietary (GPT-5, o3-mini, GPT-4o and Claude 3.7 Sonnet), open weight (DeepSeek and Qwen3) and domain-specialized medical (MedGemma) LLMs on three tasks: (i) Distinguishing tumor-somatic mutations from clonal hematopoietic (CH) variants in patients with matched tumor-whole blood profiling (N=37,179 patients; 54,807 samples), (ii) Classifying oncogenic variants using the OncoKB dataset as a positive control (N=10,489 patients; 10,752 samples; 13,470 variants), and (iii) Predicting cancer type from tumor genomic profiles in the multi-institution AACR GENIE dataset (N=97,074 patients; 102,791 samples). Results: Multiple LLMs approached the accuracy of MetaCH, a supervised model for distinguishing somatic tumor mutations from CH variants. o3-mini achieved the highest accuracy for distinguishing oncogenic driver from benign passenger mutations. Among patients with non-small cell lung cancer and mutations in KEAP1 , those with VUSs classified as oncogenic by GPT-5 had worse overall survival than those with VUSs classified as benign. GPT-5, o3-mini, and Claude 3.7 Sonnet had accuracy approaching that of a supervised model, GDD-ENS, at classifying 34 cancer types using tumor genomic profiles. Ensemble approaches combining prediction results from GPT-5 and GDD-ENS improved cross-institutional generalizability and performance by an average of 20%. In their reasoning, LLMs discussed clinically relevant genomic features consistent with feature importances from GDD-ENS. Conclusion: Without task-specific training, LLMs achieved performance comparable to specialized supervised models across all tasks. F1 score comparison across three cancer genomics prediction tasks. Mutation status (Tumor-somatic vs. CH) Oncogenic Variants (Benign vs. Oncogenic) Cancer Type (34 types) GPT-5 0.94 0.80 0.49 GPT-4o 0.96 0.75 0.32 o3-mini 0.93 0.82 0.43 Claude 3.7 Sonnet 0.95 0.79 0.43 DeepSeek 0.86 0.79 0.23 Qwen3 0.95 0.79 0.26 MedGemma 0.86 0.77 0.19 MetaCH 0.98 n/a n/a AlphaMissense n/a 0.90 n/a GDD-ENS n/a n/a 0.57

利益披露 Disclosure

J. Yu, None.. M. Darmofal, None.. M. Waters, None.. J. Choy, None. T. N. Tran, Natera Employment. C. Fu, None.. L. Morales, None.. K. U, None. R. L. Levine, Qiagen g., Board of Directors, non-salaried role), Stock. Ajax Therapeutics, Inc. g., Board of Directors, non-salaried role), Stock, ). The Mark Foundation for Cancer Research g., Board of Directors, non-salaried role). Mission Bio g., Board of Directors, non-salaried role), Stock. Kurome Therapeutics, Inc. g., Board of Directors, non-salaried role), Stock. Syndax g., Board of Directors, non-salaried role), Stock. Scorpion Therapeutics, Inc. g., Board of Directors, non-salaried role), Stock. Zentalis Pharmaceuticals g., Board of Directors, non-salaried role), Stock, ). Jubilant Therapeutics Inc. g., Board of Directors, non-salaried role), Stock. Auron Therapeutics, Inc. g., Board of Directors, non-salaried role), Stock. Prelude Therapeutics g., Board of Directors, non-salaried role), Stock. C4 Therapeutics g., Board of Directors, non-salaried role), Stock. Cure Breast Cancer Foundation ). Calico ). ECOG-ACRIN Cancer Research Group Independent Contractor. Genome Canada Independent Contractor. Goldman Sachs Independent Contractor. Astra Zeneca Independent Contractor. N. Schultz, Innovation in Cancer Informatics g., Board of Directors, non-salaried role). Stand Up to Cancer Independent Contractor. M. F. Berger, AstraZeneca Independent Contractor. Paige.AI, Inc. Independent Contractor. JCO Precision Oncology g., Board of Directors, non-salaried role). SOPHiA GENETICS S.A. g., Board of Directors, non-salaried role). Journal of Molecular Diagnostics g., Board of Directors, non-salaried role). Q. Morris, None. J. Jee, Microsoft Stock. AstraZeneca Travel.

← 返回 AACR 2026 检索