PO.BCS01.06 · 生物信息与计算
从单张H&E切片到肺肿瘤微环境中的虚拟免疫组织化学生物标志物染色
From single H&E to virtual immunohistochemical biomarker staining in the lung tumor microenvironment
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
对肿瘤微环境（TME）进行分析是理解影响癌症诊断、进展和治疗反应的免疫、基质与肿瘤相互作用的基础。物理免疫组织化学（IHC）仍然不可或缺，但受限于试剂依赖性、劳动密集型工作流程以及组织耗竭。本研究旨在开发并验证一个虚拟IHC平台，直接从苏木精-伊红（H&E）切片重现生物标志物染色，从而在保存组织的同时实现高保真的TME表征。利用ViewsML深度学习平台，训练了一个虚拟TME组合以预测五种关键生物标志物：CD31（内皮/血管系统）、CD45（全白细胞）、CD68（巨噬细胞）、平滑肌肌动蛋白/SMA（活化的成纤维细胞和周细胞）以及细胞角蛋白AE1/3（肿瘤上皮）。使用了316张数字化肺癌活检切片（Aperio GT450），涵盖超过1.5亿个标注的细胞。数据被分为训练集（70%）、验证集（15%）和测试集（15%）。神经网络针对生物标志物特异性预测进行了优化，模型性能采用ROC AUC指标和盲法病理学家评审进行评估。虚拟生物标志物与物理IHC显示出高度一致性，AUC分别为CD31=0.91、CD45=0.90、CD68=0.93、SMA=0.91和CK AE1/3=0.90。Pearson相关系数分别为0.73、0.72、0.76、0.73和0.76（p<0.0001）。虚拟染色保留了空间和形态学特征，包括肿瘤-基质边界处CD31阳性的血管框架、CD45阳性的免疫浸润、CD68阳性的巨噬细胞聚集、SMA阳性的基质反应模式以及CK阳性的肿瘤上皮。定量的单细胞生物标志物表达使得能够自动进行精确的细胞分数、细胞比率、空间邻近度以及免疫和巨噬细胞聚集分析。这些发现表明，ViewsML的虚拟IHC技术能够从标准H&E切片重现物理染色，为数字化TME分析提供了一种可扩展的解决方案并保存组织。该方法通过实现免疫主导型与基质主导型肿瘤的分层以指导免疫治疗，支持转化研究、药物开发中的虚拟生物标志物筛选以及临床应用。未来的工作将把该平台扩展到更多肿瘤类型，并纳入多重虚拟染色，以实现更广泛的数字病理学整合。
查看英文原文 English abstract
Profiling the tumor microenvironment (TME) is fundamental to understanding immune, stromal, and tumor interactions influencing cancer diagnosis, progression, and therapeutic response. Physical immunohistochemistry (IHC) remains essential but is limited by reagent dependency, labor-intensive workflows, and tissue exhaustion. This study aimed to develop and validate a virtual IHC platform to reproduce biomarker staining directly from hematoxylin and eosin (H&E) slides, enabling high-fidelity TME characterization while conserving tissue. Using the ViewsML deep learning platform, a virtual TME panel was trained to predict five key biomarkers: CD31 (endothelial/vasculature), CD45 (pan-leukocyte), CD68 (macrophage), smooth muscle actin/SMA (activated fibroblasts and pericytes), and cytokeratin AE1/3 (tumor epithelium). 316 digitized lung cancer biopsy slides (Aperio GT450) encompassing over 150 million annotated cells were used. Data were split into training (70%), validation (15%), and test (15%) sets. Neural networks were optimized for biomarker-specific predictions, and model performance was evaluated using ROC AUC metrics and blinded pathologist review. Virtual biomarkers showed strong concordance with physical IHC, achieving AUCs of CD31=0.91, CD45=0.90, CD68=0.93, SMA=0.91, and CK AE1/3=0.90. Pearson's correlations were 0.73, 0.72, 0.76, 0.73, and 0.76 (p<0.0001). Virtual stains preserved spatial and morphological features, including CD31-positive vascular frameworks at tumor-stroma boundaries, CD45-positive immune infiltration, CD68-positive macrophage aggregates, SMA-positive stromal reaction patterns, and CK-positive tumor epithelium. Quantitative per-cell biomarker expression enabled automated precise cell fraction, cell ratio, spatial proximity, and immune and macrophage clustering analysis. These findings demonstrate that ViewsML's virtual IHC technology can reproduce physical staining from standard H&E slides, offering a scalable solution for digital TME profiling and conserving tissue. This approach supports translational research, virtual biomarker screening in drug development, and clinical application by enabling immune versus stroma-dominant tumor stratification to guide immunotherapy. Future work will extend this platform to more tumor types and incorporate multiplex virtual staining for broader digital pathology integration.
利益披露 Disclosure
K. To, None..
C. Jackson, None.
L. Vaickus,
ViewsML Stock, Stock Option.
L. Schobs, None..
R. Kamra Lyons, None..
R. Azhar, None.