Integrating cancer genomics and spatial architecture of tumor infiltrating lymphocytes

摘要 肿瘤浸润性淋巴细胞（TIL）是驻留在肿瘤中的免疫细胞的重要组分 微环境（TME）。TME中TIL的类型和数量对总生存率有影响， 对免疫疗法的反应的指标。尽管它们作为患者免疫功能的指标很重要， 针对癌症的反应，分析来自大型人群数据集的TILS存在多种挑战， 数以千计的样本。缺乏能够自动分析组织病理学图像以用于诊断的方法。 不同的特征，如TIL的空间分布，它们与相邻细胞的拓扑相互作用 及其与特定临床结局的关系。更具挑战性的是集成TIL指标 with cancer癌症genome基因组data数据.大多数其他方法提供了TIL的定性指标，并且经常依赖于人工 来自病理学家的检查-这种方法缺乏可扩展性并且受到观察者偏差的影响。解决这些 挑战，我们开发了一个计算框架，使用深度学习模型来识别多个细胞 组织病理学图像的类型。我们方法的主要创新是分子标记转移， 注释从组织病理学图像中提取的数万个小区域，而无需手动检查。 这种方法非常准确，高效，可扩展，易于自动化，可用于分析数百万张图像。 该项目的目标是解决深度学习应用中的关键挑战， 组织病理学图像：大量标记图像作为训练数据集。我们有三个具体目标：1） 从癌症基因组图谱的10，000多张组织病理学图像中确定TIL的空间定量 项目; 2）将TIL度量与克隆肿瘤突变负荷（TMB）相关联; 3）确定TIL与克隆肿瘤突变负荷（TMB）的相关性。 免疫检查点阻断反应。这项研究是重要的，因为我们的方法使一个 在细胞水平上从组织病理学图像中全面表征TIL，使用的数据是 通常在临床环境中可获得，并且可以容易地与癌症基因组数据整合。

ABSTRACT Tumor infiltrating lymphocytes (TILs) are an important component of the immune cells that reside in the tumor microenvironment (TME). The type and number of TILs in the TME have an impact on overall survival and are an indicator of response to immunotherapy. Despite their importance as an indicator of a patient’s immune response to cancer, there are multiple challenges for analyzing TILS from large population data sets involving thousands of samples. There is a lack of methods that can automate an analysis of histopathologic images for different features such as the spatial distribution of TILs, their topological interactions with their neighboring cells in the TME and their association with specific clinical outcomes. Even more challenging is integrating TIL metrics with cancer genomic data. Most other methods provide qualitive metrics of TILs and frequently rely on manual inspection from pathologists – this approach lacks scalability and is subject to observer bias. To address these challenges, we developed a computational framework that uses a deep learning model to identify multiple cell types from histopathology images. The major innovation of our approach is molecular label transferring that annotates tens of thousands of small areas extracted from histopathology images without manual inspections. This approach is highly accurate, efficient, scalable and readily automated for the analysis of millions of images. The objective of this project is to address a key challenge in the application of deep learning to histopathological image: large number of labeled images as training data set. We have three specific aims to 1) identify spatial quantification of TILs from over 10,000 histopathological images from the Cancer Genome Atlas Project; 2) correlate TIL metrics with clonal tumor mutation burden (TMB); 3) determine association of TILs with immune checkpoint blockade responses. This research is significant because our approach enables for a comprehensive characterization of TILs from histopathological images at cellular level, using data that is commonly accessible in clinical settings and can be readily integrated with cancer genomic data.