PO.BCS01.07 · 生物信息与计算

深度学习整合分子与组织病理学数据用于非小细胞肺癌的预后分层

Deep learning integration of molecular and histopathological data for prognostic stratification in non small cell lung cancer

编号 1448 展板 11 时间 4/20 09:00–12:00 区域 Section 4 主讲 Sanddhya Jayabalan, BS;MS

分会场 Digital Pathology 2

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

Sanddhya Jayabalan¹, Konstantinos Efthymiadis², Alexia Eliades³, Kyriaki Papadopoulou², Abraham Pouliakis⁴, Elena Fountzilas², Sofia Lampaki², Mattheos Bobos², Anna Goussia², Soultana Meditskou², Konstantinos Kyritsis⁵, Helena Linardou², George Pentheroudakis², Dimitrios Bafaloukos², Dimitrios Pectasides², Epaminondas Samantas², Zunamys I. Carrero¹, George Fountzilas², Jakob N. Kather¹

¹Else Kröner Fresenius Center for Digital Health (EKFZ), Technische Universität Dresden, Dresden, Germany,²Hellenic Cooperative Oncology Group (HeCOG), Athens, Greece,³Medicover Genetics, Cyprus, Greece,⁴University General Hospital Attikon, National and Kapodistrian University of Athens Medical School, Athens, Greece,⁵Centre for Research and Technology Hellas, Institute of Applied Biosciences, Thessaloniki, Greece

摘要 Abstract

中文摘要

背景：共突变、PD-L1和TILs是关键的NSCLC生物标志物。我们将深度学习应用于多模态患者队列，以识别整合形态学、突变和临床特征的预后模式。方法：对来自18个希腊肿瘤协作组附属中心的367名NSCLC患者进行回顾性评估，检测PD-L1状态(Dako 22C3 pharmDx)、TILs(H&E切片)以及使用38基因二代测序(NGS)组合的体细胞致病性变异。全切片图像(WSI)通过光学显微镜扫描仪进行数字化。对十个病理学基础模型进行基准测试，以预测突变、共突变、PD-L1和TILs状态。流行率>5%的突变基因被纳入突变和共突变终点考量(TP53、KRAS、STK11、PTEN、EGFR)。在WSI特征上训练了一个视觉transformer模型来预测终点并评估AUROC。Kaplan-Meier分析评估了模型的预后相关性，模型注意力图中的顶部特征瓦片提供了形态学可解释性。STAMP数字病理学流程支持特征提取和模型训练。结果：单突变模型的AUROC评分为0.6-0.85，其中从HOptimus1特征预测STK11最高。共突变模型产生0.69-0.77的AUROC评分，其中从Uni2特征预测EGFR-TP53最佳。KRAS-TP53共突变模型(AUROC 0.69，Uni2)在各类别间的总生存曲线中显示出显著分离(p=0.05)。表现最佳的PD-L1和TIL模型也显示出显著的生存分离(p=0.005和p=0.05)。结论：研究结果证明了病理学基础模型在为NSCLC构建具有多模态可解释性的复杂临床相关预后模型方面的潜力。按终点和基础模型的AUROC TP53 KRAS PTEN STK11 EGFR EGFR+TP53 KRAS+STK11 KRAS+TP53 PTEN+TP53 TILs PD-L1 Conch1.5 0.48 0.62 0.24 0.79 0.61 0.58 0.72 0.26 0.41 0.65 0.49 CTranspath 0.52 0.77 0.45 0.17 0.5 0.57 0.33 0.68 0.34 0.6 0.6 Dinobloom 0.43 0.45 0.51 0.58 0.46 0.75 0.75 0.68 0.38 0.65 0.73 Gigapath 0.47 0.5 0.44 0.84 0.31 0.61 0.37 0.12 0.63 0.65 0.6 HOptimus0 0.5 0.72 0.56 0.65 0.55 0.61 0.63 0.08 0.48 0.62 0.65 HOptimus1 0.57 0.66 0.43 0.85 0.64 0.63 0.52 0.39 0.28 0.61 0.55 Musk 0.49 0.49 0.51 0.77 0.54 0.61 0.53 0.58 0.43 0.7 0.54 Plip 0.55 0.52 0.5 0.52 0.43 0.48 0.72 0.44 0.74 0.67 0.6 Uni 0.61 0.58 0.6 0.56 0.48 0.7 0.45 0.26 0.46 0.66 0.61 Uni2 0.52 0.69 0.51 0.81 0.61 0.77 0.41 0.69 0.57 0.59 0.64

查看英文原文 English abstract

Background: Co-mutations, PD-L1 and TILs are key NSCLC biomarkers. We applied deep learning to a multimodal patient cohort to identify prognostic patterns integrating morphology, mutations, and clinical features. Methods: 367 NSCLC patients from 18 Hellenic Cooperative Oncology Group-affiliated centers were retrospectively assessed for PD-L1 status (Dako 22C3 pharmDx), TILs (H&E slides), and somatic pathogenic variants with a 38-gene next-generation sequencing (NGS) panel. Whole slide images (WSI) were digitized by an optical microscope scanner. Ten pathology foundation models were benchmarked for predicting mutation, co-mutation, PD-L1 and TILs status. Mutated genes with >5% prevalence were considered for mutation and co-mutation endpoints ( TP53 , KRAS , STK11 , PTEN , EGFR ). A vision transformer model was trained on WSI features to predict endpoints and evaluate AUROC. Kaplan-Meier analysis assessed prognostic relevance of models and top feature tiles from model attention maps provided morphological explainability. The STAMP digital pathology pipeline supported feature extraction and model training. Results: Single mutation models yielded AUROC scores of 0.6-0.85, with STK11 prediction from HOptimus1 features highest. Co-mutation models produced AUROC scores of 0.69-0.77 with EGFR-TP53 prediction from Uni2 features the best. The KRAS-TP53 co-mutation model (AUROC 0.69, Uni2) showed significant separation in overall survival curves (p=0.05) between classes. Best-performing PD-L1 and TIL models also demonstrated significant survival separation (p=0.005 and p=0.05). Conclusion: Findings demonstrate the potential of pathology foundation models to derive complex clinically-relevant prognostic models for NSCLC with multimodal explainability. AUROC by endpoint and foundation model TP53 KRAS PTEN STK11 EGFR EGFR+TP53 KRAS+STK11 KRAS+TP53 PTEN+TP53 TILs PD-L1 Conch1.5 0.48 0.62 0.24 0.79 0.61 0.58 0.72 0.26 0.41 0.65 0.49 CTranspath 0.52 0.77 0.45 0.17 0.5 0.57 0.33 0.68 0.34 0.6 0.6 Dinobloom 0.43 0.45 0.51 0.58 0.46 0.75 0.75 0.68 0.38 0.65 0.73 Gigapath 0.47 0.5 0.44 0.84 0.31 0.61 0.37 0.12 0.63 0.65 0.6 HOptimus0 0.5 0.72 0.56 0.65 0.55 0.61 0.63 0.08 0.48 0.62 0.65 HOptimus1 0.57 0.66 0.43 0.85 0.64 0.63 0.52 0.39 0.28 0.61 0.55 Musk 0.49 0.49 0.51 0.77 0.54 0.61 0.53 0.58 0.43 0.7 0.54 Plip 0.55 0.52 0.5 0.52 0.43 0.48 0.72 0.44 0.74 0.67 0.6 Uni 0.61 0.58 0.6 0.56 0.48 0.7 0.45 0.26 0.46 0.66 0.61 Uni2 0.52 0.69 0.51 0.81 0.61 0.77 0.41 0.69 0.57 0.59 0.64

利益披露 Disclosure

S. Jayabalan, AstraZeneca Employment. K. Efthymiadis, HeSMO (Hellenic Society of Medical Oncology) ). A. Eliades, None.. K. Papadopoulou, None.. A. Pouliakis, None.. E. Fountzilas, None.. S. Lampaki, None.. M. Bobos, None.. A. Goussia, None.. S. Meditskou, None.. K. Kyritsis, None.. H. Linardou, None.. G. Pentheroudakis, None.. D. Bafaloukos, None.. D. Pectasides, None.. E. Samantas, None.. Z. I. Carrero, None.. G. Fountzilas, None. J. N. Kather, AstraZeneca ), Other, Consulting. Panakeia Other. Bioptimus Other, Consulting. StratifAI Stock. Synagen Stock. Spira Labs Stock. GSK ).

← 返回 AACR 2026 检索