PO.BCS01.08 · 生物信息与计算
利用深度学习从H&E切片开发虚拟Cyclin E1生物标志物以预测妇科恶性肿瘤中Cyclin E1的过表达
Development of a virtual Cyclin E1 biomarker using Deep Learning from H&E slides for predicting Cyclin E1 overexpression in gynecological malignancy
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
本研究的目的是开发并验证一种虚拟免疫组化（vIHC）算法，该算法能够从苏木精和伊红（H&E）染色的全切片图像中预测妇科恶性肿瘤（主要涵盖高级别浆液性卵巢癌（HGSOC）和子宫浆液性癌（USC））中的Cyclin E1（CCNE1）蛋白表达。CCNE1是一种关键的细胞周期调节因子，其基因扩增（拷贝数≥6）与蛋白过表达（H-score >50）以及对WEE1抑制剂敏感性增强密切相关。传统IHC需要宝贵的组织和额外的实验室处理；源自常规可获得的H&E切片的ViewsML虚拟生物标志物平台提供了一种可扩展的替代方案，以加速靶向治疗的患者筛选。
H&E及配对的Cyclin E1 IHC全切片图像由Debiopharm提供。分析了60对数字切片（40例HGSOC和20例USC）。数据集被分为训练（n=45）、验证（n=6）和测试（n=9）队列。ViewsML利用经训练的神经网络模型来学习可预测Cyclin E1表达强度（0-3+）的形态学和核特征。使用每个细胞的敏感性、特异性以及与物理IHC强度和H-score分类的一致性（包括ROC AUC指标）来评估模型性能。通过区分弱、中、强染色模式的核定位来评估预测的和真实的Cyclin E1表达之间的一致性，从而能够定量评估HGSOC和USC中Cyclin E1阳性肿瘤比例。
总之，本研究证明了针对Cyclin E1的AI驱动虚拟生物标志物的可行性，它能够直接从H&E切片预测蛋白过表达。这种虚拟IHC方法节省了宝贵的组织，并加速了用于患者筛选的生物标志物筛查，有助于改善对Cyclin E1相关治疗试验的富集。未来的应用包括与多重虚拟标志物整合，以进一步增强临床适用性。
查看英文原文 English abstract
The purpose of this study was to develop and validate a virtual immunohistochemistry (vIHC) algorithm capable of predicting Cyclin E1 (CCNE1) protein expression from hematoxylin and eosin (H&E)-stained whole-slide images in gynecological malignancy encompassing primarily high-grade serous ovarian carcinoma (HGSOC) and uterine serous carcinoma (USC). CCNE1 is a key cell-cycle regulator whose gene amplification (copy number ≥6) correlates strongly with protein overexpression (H-score >50) and enhanced sensitivity to WEE1 inhibitors. Conventional IHC requires precious tissue and additional laboratory processing; the ViewsML virtual biomarker platform derived from routinely available H&E slides provides a scalable alternative to accelerate patient selection for targeted therapy.
H&E and matched Cyclin E1 IHC whole-slide images were provided by Debiopharm. Sixty paired digital slides (40 HGSOC and 20 USC) were analyzed. The dataset was divided into training (n=45), validation (n=6), and testing (n=9) cohorts. ViewsML utilized neural network models trained to learn morphological and nuclear features predictive of Cyclin E1 expression intensity (0-3+). Model performance was evaluated using per-cell sensitivity, specificity, and concordance with physical IHC intensity and H-score classifications, including ROC AUC metrics. Concordance between predicted and true Cyclin E1 expression was evaluated through nuclear localization distinguishing weak, moderate, and strong staining patterns, allowing quantitative assessment of Cyclin E1-positive tumor fractions across HGSOC and USC.
In conclusion, this study demonstrates the feasibility of an AI-driven virtual biomarker for Cyclin E1 that can predict protein overexpression directly from H&E slides. The virtual IHC approach conserves valuable tissue and accelerates biomarker screening for patient selection, facilitating improved enrichment for Cyclin E1-associated therapeutic trials. Future applications include integration with multiplex virtual markers to further enhance clinical applicability.
利益披露 Disclosure
J. Fuchs,
Debiopharm International Employment.
K. To, None..
C. Jackson, None..
L. Schobs, None..
R. Lyons, None..
R. Azhar, None.