PO.TB10.07 · 肿瘤生物学
AI引导的空间多组学整合揭示前列腺癌人群差异背后的免疫与基质异质性。
AI-guided spatial multi-omics integration reveals immune and stromal heterogeneity underlying population disparities in prostate cancer.
该海报暂无可下载的资料
AACR 官方页面
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
前列腺癌对非裔美国（AA）男性的影响尤为突出，他们的发病率和死亡率均高于欧裔美国（EA）男性。本研究旨在开发并应用一种基于人工智能（AI）的框架，通过对肿瘤微环境（TME）进行空间表征，探讨造成这些差异的生物学机制。我们采用了CODA这一计算方法，该方法能够从连续切片的苏木精-伊红（H&E）染色玻片中以亚细胞分辨率重建大块组织，用于组织学分割和多组学整合。CODA使用来自AA和EA患者的600张H&E染色前列腺肿瘤切片进行训练，采用了一种针对九个区室（包括肿瘤、基质、脉管系统和炎症）像素级分类进行优化的深度卷积神经网络。使用混淆矩阵评估的模型性能达到了超过90%的分类准确率，并在各队列中稳健地实现了泛化。CODA分割图谱与10x Genomics Visium HD空间转录组数据进行了对齐，以生成具有生物学注释的区域，例如腔面分泌区、基质区、炎症上皮区和免疫区。功能分析揭示了不同的TME生态位，包括适应性免疫、髓系富集、细胞毒性、成纤维细胞和腔面上皮区室。AA肿瘤表现出比EA肿瘤更高的免疫浸润和炎症活性，特征为致密的淋巴样聚集体和细胞因子信号富集。与来自AKOYA IO60面板（60个免疫和基质标志物）的空间蛋白质组学数据整合，证实了免疫簇（CD20+、CD3e+、CD68+、CD8+），并揭示了转录本表达与蛋白信号之间存在一定程度的不一致，提示存在情境依赖性调控。来自连续切片肿瘤（一例AA、一例EA）的三维重建可视化了区室连续性以及免疫-基质界面。总的来说，这些发现提供了人群特异性免疫与基质异质性的新证据，这些异质性可能是造成前列腺癌差异的原因，并展示了AI引导的空间多组学在推进公平精准肿瘤学方面的效用。
查看英文原文 English abstract
Prostate cancer disproportionately affects African American (AA) men, who experience higher incidence and mortality rates than European American (EA) men. The purpose of this study was to develop and apply an artificial intelligence (AI)-based framework to investigate biological mechanisms contributing to these disparities through spatial characterization of the tumor microenvironment (TME). We employed CODA, a computational method that reconstructs large tissues at subcellular resolution from serially sectioned hematoxylin and eosin (H&E)-stained slides, for histological segmentation and multi-omics integration.CODA was trained on 600 H&E-stained prostate tumor sections from AA and EA patients using a deep convolutional neural network optimized for pixel-level classification of nine compartments, including tumor, stroma, vasculature, and inflammation. Model performance, evaluated using a confusion matrix, achieved over 90% classification accuracy and generalized robustly across cohorts. CODA segmentation maps were aligned with 10x Genomics Visium HD spatial transcriptomic data to generate biologically annotated domains such as luminal secretory, stromal, inflammatory epithelial, and immune regions.Functional analysis revealed distinct TME niches, including adaptive immune, myeloid-rich, cytotoxic, fibroblast, and luminal-epithelial compartments. AA tumors demonstrated higher immune infiltration and inflammatory activity than EA tumors, characterized by dense lymphoid aggregates and cytokine signaling enrichment. Integration with spatial proteomics data from the AKOYA IO60 panel (60 immune and stromal markers) confirmed immune clusters (CD20+, CD3e+, CD68+, CD8+) and revealed a level of discordance between transcript expression and protein signaling, suggesting context-dependent regulation.Three-dimensional reconstructions from serially sectioned tumors (one AA, one EA) visualized compartment continuity and immune-stromal interfaces. Collectively, these findings provide new evidence of population-specific immune and stromal heterogeneity that may contribute to prostate cancer disparities and demonstrate the utility of AI-guided spatial multi-omics for advancing equitable precision oncology.
利益披露 Disclosure
S. M. Mwamburi, None..
J. Layne, None..
N. Talebian, None..
E. G. Baraban, None..
A. Kiemen, None..
C. C. Yates, None.