Using Cyclic Immunofluorescence (CyCIF), Deep-learning, and Human Tumor Microarrays (hTMAs) for Antibody Validation and Creation of an Atlas of Cancers Emphasizing Unique Phenotypes

Abstract

All cancers are different from each other, in their mutation spectrum, primary location of occurrence, and metastases as well as their phenotypic traits and tumor composition. For years, clinicians and pathologists have sought to describe, characterize, and exploit the hallmarks of cancer broadly to identify specific differences between cancer types and subtypes. As medical technology and science advances, so has understanding of the underlying biology of cancer initiation and progression. Further, the ability to identify and visualize subtle differences between types of cancer based on their phenotypic and morphologic presentation aids in the research and design of newer targeted cancer therapies aimed to improve the lives of patients regardless of their original prognosis. Currently, the most widely used method for studying tumor tissue and diagnosis diseases is hematoxylin and eosin staining (H&E) of surgical biopsies with supplemental immunohistochemistry (IHC) or in-situ hybridization (ISH) for molecular biomarkers. These methods are low-plex but highly reliable and have remained the gold standards half a century. IHC typically identifies and quantifies one biomarker of interest at a time and answer specific questions, such as whether or not a particular breast tumor biopsy contains cells that overexpress human epidermal growth factor receptor 2 (HER2) or not. IHC stains require manual interpretation by trained pathologists or histologists based on counting cells (commonly with a binary call as either positive or negative) expressing a particular marker: if there is an abundance of brown-stained HER2+ cells scored on one patient’s biopsy slide as compared to a control, then HER2 is scored as over-expressed. Unfortunately, this is a slow, laborious process that does not take advantage of many advances in computer vision and image analysis. Thus, digital methods that incorporate multiple antibodies have long been sought. Recently, a modified version of the traditional immunofluorescence assay, called cyclic immunofluorescence (CyCIF), has been used to perform high-throughput, multiplexed imaging of fixed, cultured cells in a multiwell plate format and (Lin et al., 2015) as well as formalin-fixed paraffin-embedded (FFPE) sections of tissue or tumor samples (Lin et al., 2018) on slides in a manner compatible with standard histopathology workflows. As with the IHC assays, the goal is to detect particular cell- or tissue-specific macromolecules with antibodies, but for more analytes to spare tissue and reveal deeper cellular phenotypes. In CyCIF, both unconjugated and conjugated commercially available antibodies, used for indirect IF and direct IF, respectively, can be used, with up to 60 or more different antigens measured on each specimen. This enables a thorough characterization of tissue-intrinsic properties as well as tumor. In doing so, we can provide a more comprehensive and specific understanding of tissue sections and cellular populations as opposed to simply the presence or absence of an antigen or disease. Whole slide imaging (WSI) is performed on tumor biopsies, cycles are registered and stitched together, and a mutli-channel images containing data from numerous antibody stains is then analyzed to define the biological context in which cancer continues to persist. Here, we used CYCIF to investigate the composition and architecture of human cancers and adjacent normal tissue as a means to describe cellular composition both inside and outside of the tumor and along the tumor boundary, explore cell-to-cell interactions and neighborhood analyses, as well as evaluate the efficacy of commercially available antibodies. The data we present here, as well as the EMIT dataset on which it is based (Schapiro et al., 2021), defines, characterizes, and highlights noticeable trends of biology in human tissue microarrays (hTMA). I will use the staining of complementary serial sections of TMAs to accomplish our aim of characterizing tumor and benign cell morphology across cancers, uncover patterns across cancer different cancer types, identify morphological distinctions through antibody expression, and evaluate antibody staining across hundreds of reagents through a multi-disciplinary approach. This, beginning with fine-tuned and well-controlled experimentation and the use of deep learning artificial intelligence (AI) and automated post-processing of image-based data. In achieving our aims and showcasing the potential of CyCIF, we hope to translate this current research into an atlas-like resource aimed at addressing many outstanding questions in the fields of histology, pathology, and the newly emerging field of digital histopathology. Though we do not CyCIF does not provide a direct link into clinical care, we show that the results produced through our multifaceted pipeline can, and should, have a direct impact on clinical follow-ups as they pertain to individualized medicine. In detailing a patient’s primary tumor biopsy in replicate, with uniquely built-upon CyCIF panels to unearth as much biological and immunological information as possible, we have created an argument as to why high-plex cancer atlases might serve as valuable tools for enhancing our knowledge of disease-specific underlying biology to a never- before-possible standard.