The single-cell pathology landscape of breast cancer

Abstract

Single-cell analyses have revealed extensive intra-and inter-patient cancer heterogeneity 1 , but complex single-cell phenotypes and their spatial context are not yet reflected in the histologic stratification that is the foundation of many clinical decisions. Here, we used imaging mass cytometry 2 to simultaneously quantify 35 biomarkers resulting in 720 highdimension immunohistochemistry pathology images of tumor tissue from 352 breast cancer patients for whom long-term survival data were available. Spatial, single-cell analysis identified tumor and stromal single-cell phenotypes, their organization and heterogeneity, and enabled categorization of breast cancer cellular architecture based on cellular composition and tissue organization. Our analysis revealed multi-cellular features of the tumor microenvironment and novel breast cancer subgroups associated with distinct clinical outcomes. Thus, spatially resolved, single-cell analysis can characterize intra-tumor phenotypic heterogeneity in a disease-relevant manner with the potential to inform patient-specific diagnosis. MainHistologic and phenotypic differences between tumors guide cancer diagnosis, prognosis, and selection of treatment. Currently, breast cancer patients are graded based on tumor structure and cellular morphology, and subcategorized when more than 1% of tumor cells contain hormone receptors or more than 10% express high levels of HER2 or have amplified HER2 [3][4][5] . This leaves a large portion of cells uncharacterized even though additional molecular subclasses and morphologic features have been identified as prognostic [6][7][8][9] . It is clear that clonal evolution and spatially distinct tumor microenvironments drive inter-and intra-patient cellular

Channel Embedding for Informative Protein Identification from Highly Multiplexed Images

“…for assessing tumor grade. Moreover, the informative channels identified by our novel method align with findings from a single cell data analysis [5], even though our approach does not require single cell segmentation.…”

“…To evaluate informative channel identification performance, we define a task in which modern deep neural networks [4,20] achieve high accuracy. For a real-world application, we apply our model to a task of predicting tumor grade from a breast cancer dataset generated using IMC [5].…”

“…We synthesize 600 training, 300 validation, and 300 test images with 30 channels. Figure 4 visualizes three randomly-selected channels of a synthetic image compared to a real highly multiplexed image [5]. Evaluation Metric: To evaluate classification performance, we measure accuracy, which is the number of correct predictions divided by the number of total images.…”

“…As such, highly multiplexed imaging is emerging as a potentially breakthrough technology in translational research and clinical diagnosis. Examples of highly multiplexed imaging technologies include imaging mass cytometry (IMC) [5], multiplexed ion beam imaging (MIBI) [6], co-detection by indexing (CODEX) [3], and cyclic immunofluorescence (CyCIF) [8].…”

“…The process is illustrated in Figure 1. What follows is a description of our system for the goal of identifying the most informative channels for assessing the tumor grade of highly multiplexed images [5]. We first encode each channel using the shared weights of a ResNet18 [4] backbone encoder.…”