Deep-learning Integration of Histopathology and Proteogenomics at a Pan-cancer Level - Resubmission

项目总结 发现和理解与蛋白质组学相关的新的病理特征在 改善和简化癌症预后和治疗。目前，生物医学研究正在努力预测 癌症的结果取决于在临床环境中不易获得的测序方法。相反，临床医生 经常依靠组织病理学图像来目测组织形态的异常变化。 随后，一种直接从组织病理学图像推断临床和分子特征的工具将被利用 组学研究的力量与基于图像的诊断的可行性。 在寻求更好地了解肿瘤生物学方面，组织病理学成像数据仍然远远没有得到充分利用，主要是 因为分析和数据整合的工具不足。确定泛癌的特征和保守的 肿瘤发生的途径，因此我建议开发多分辨率深度卷积神经网络 (CNN)10种不同癌症类型的模型，并预测临床注释、组织学结果和危急情况 基于肿瘤组织病理学图像的突变(目标1)。通过我们的实验室与国家实验室的合作 癌症研究所的临床蛋白质组肿瘤分析联盟(CPTAC)，我们可以访问多组学， 来自1,602名患者的临床和组织病理学数据。此外，要了解生物 驱动形态变化的机制，我建议开发计算方法，将 具有成像功能的转录和蛋白质组表达数据集，以促进通道级知识发现 (目标2)。我们的建议是第一个将表达扰动与形态模式相关联并识别 直接从组织病理学图像中丰富典型通路。重要的是，这项提案旨在将 利用临床医生的诊断工具进行生物医学研究的科学努力，以扩大诊断能力和 改进临床实践。

PROJECT SUMMARY Discovering and understanding novel pathological features that correlate with proteogenomics is crucial in improving and streamlining cancer prognosis and treatment. Currently, biomedical research efforts to predict cancer outcomes rely on sequencing approaches not readily accessible in a clinical setting. Instead, clinicians frequently rely on histopathology images to visually assess for aberrant changes to tissue morphology. Subsequently, a tool to infer clinical and molecular signatures directly from histopathology images would harness the power of omics research with the feasibility of image-based diagnosis. Histopathology imaging data is still vastly underutilized in the quest to better understand tumor biology, largely because of inadequate tools for analysis and data integration. To identify pan-cancer hallmarks and conserved pathways of tumorigenesis, I therefore propose to develop multi-resolution deep convolutional neural network (CNN) models across 10 different cancer types and predict clinical annotations, histology outcomes, and critical mutations based on tumor histopathology images (Aim 1). Through our lab’s collaboration with the National Cancer Institute’s Clinical Proteomic Tumor Analysis Consortium (CPTAC), we have access to multi-omics, clinical, and histopathologic data obtained from 1,602 patients. In addition, to understand the biological mechanisms driving morphology changes, I propose to develop computational methods that integrate transcriptomic and proteomic expression datasets with imaging to facilitate pathway-level knowledge discovery (Aim 2). Our proposal is the first to correlate expression perturbations with morphology patterns and identify enriched canonical pathways directly from histopathology images. Importantly, this proposal aims to connect scientific efforts of biomedical research with the diagnostic tools of clinicians to expand diagnostic power and improve clinical practice.