PO.BCS01.14 · 生物信息与计算

一种基于RNA的生存模型预测真实世界中对trastuzumab deruxtecan的反应

An RNA-based survival model predicting real-world response to trastuzumab deruxtecan

海报缩略图:一种基于RNA的生存模型预测真实世界中对trastuzumab deruxtecan的反应
编号 6883 展板 27 时间 4/22 09:00–12:00 区域 Section 3 主讲 Josh Wheeler, MD;PhD
分会场 Network Biology and Precision Medicine
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

Klemen Žiberna, Anže Lovše, Žan Kuralt, Janez Kokošar, Marcel Levstek, Luka Ausec, Miha Štajdohar, Rafael Rosengarten, Mark Uhlik, Joshua Wheeler

Genialis, Inc., Boston, MA

摘要 Abstract

中文摘要
抗体药物偶联物(ADC),如trastuzumab deruxtecan(T-DXd,Enhertu),已重新定义了HER2表达型乳腺癌的治疗,然而在HER2阳性、低表达和超低表达疾病中,临床获益仍难以预测。目前的IHC/FISH诊断可量化受体丰度,却无法捕捉决定ADC敏感性的分子状态。为弥补这一空白,我们利用Genialis Supermodel(一个大型分子基础模型)为Enhertu开发了一种基于RNA的生存模型。Supermodel将基因表达映射到数百个生物模块(biomodules)中,这些生物模块是生物学的算法表征,能够捕捉多种肿瘤学特征,包括信号通路、应激反应和药物-靶点机制。我们将ADC作用机制特异性的生物模块作为预测模型的输入特征,以学习与T-DXd反应相关的生物学模式。在来自Tempus真实世界多模态数据库的一个真实世界临床队列(n≈90例接受T-DXd治疗的患者)中,我们对至下一次治疗的时间(rwTTNT)进行了生存建模。采用分层嵌套交叉验证来评估模型的稳健性和预测性能。在既往治疗线的rwTTNT以及一个独立的临床匹配队列中评估了预后特异性。该模型显示出具有统计学意义的区分能力(C指数0.632,HR 2.22 [95% CI 1.14-4.35],p = 0.017)。预测获益的患者具有更长的rwTTNT(345天对245天),且在对照队列中未出现预后信号(C指数≈0.5),提示具有预测特异性。最重要的预测特征与ADC生物学相符,包括TOP3B和TOP2A(拓扑异构酶载荷)、ATM和TP53(DNA损伤反应)、HIF1A(缺氧)、ESR1(激素信号传导)以及XBP1/NFATC1(应激和免疫调节)。这一Enhertu生存模型将生物学结构化的AI应用于真实世界RNA-seq数据,以揭示治疗特异性的反应模式。通过整合大规模嵌入、机制性生物模块和生存建模,我们识别出与T-DXd获益相关的、涉及DNA修复和应激反应的生物学程序。
查看英文原文 English abstract
Antibody-drug conjugates (ADCs) such as trastuzumab deruxtecan (T-DXd, Enhertu) have redefined therapy for HER2-expressing breast cancer, yet clinical benefit remains unpredictable across HER2-positive, -low, and -ultralow disease. Current IHC/FISH diagnostics quantify receptor abundance but fail to capture the molecular state that governs ADC sensitivity.To address this gap, we developed an RNA-based survival model for Enhertu using the Genialis Supermodel, a large molecular foundation model. The Supermodel maps gene expression into hundreds of biomodules, algorithmic representations of biology that capture diverse oncologic hallmarks including signaling pathways, stress responses, and drug-target mechanisms. We used biomodules specific to ADC mechanisms-of-action as input features in predictive models that learn biological patterns associated with T-DXd response.In a real-world clinical cohort (n≈90 T-DXd-treated patients) from the Tempus real-world multimodal database, we performed survival modeling of time-to-next-treatment (rwTTNT). Stratified nested cross-validation was used to assess model robustness and predictive performance. Prognostic specificity was assessed in prior-line rwTTNT and in an independent clinically matched cohort. The model showed statistically significant discrimination (C-index 0.632, HR 2.22 [95% CI 1.14-4.35], p = 0.017). Predicted-benefit patients had longer rwTTNT (345 vs 245 days), and no prognostic signal appeared in control cohorts (C-index ≈ 0.5), suggesting predictive specificity. Top predictive features aligned with ADC biology, including TOP3B and TOP2A (topoisomerase payload), ATM and TP53 (DNA damage response), HIF1A (hypoxia), ESR1 (hormone signaling), and XBP1/NFATC1 (stress and immune regulation).This Enhertu survival model applies biologically structured AI to real-world RNA-seq data to reveal treatment-specific patterns of response. Integrating large-scale embeddings, mechanistic biomodules, and survival modeling, we identified biological programs related to DNA repair and stress response associated with T-DXd benefit.
利益披露 Disclosure
K. Žiberna, Genialis, Inc Employment. A. Lovše, Genialis, Inc Employment. Ž. Kuralt, Genialis, Inc Employment. J. Kokošar, Genialis, Inc Employment. M. Levstek, Genialis, Inc Employment. L. Ausec, Genialis, Inc Employment. M. Štajdohar, Genialis, Inc Employment. R. Rosengarten, Genialis, Inc Employment. M. Uhlik, Genialis, Inc Employment. J. Wheeler, Genialis, Inc Employment.

← 返回 AACR 2026 检索