Cancer tissue atlases provide valuable resources for the analysis of cell differentiation, microenvironment composition, tissue architecture, and spatial relationships in 2D, 3D, and hyper-dimensional reference frames. Highly multiplexed immunofluorescence methods such as t-CyCIF (Lin, Izar, Wang, Yapp, Mei, P. Shah, et al., 2018) provides a highly customizable and readily applicable methodology to generate large-scale datasets amenable to image-based computational and visual analysis from macroscopic to sub-cellular resolution. However, few such data sets have been generated or rigorously annotated to date.

Here, we present a virtual atlas of a representative 1.6cm x 1.6cm x 5um formalin-fixed paraffin-embedded (FFPE) section of a primary colorectal adenocarcinoma subjected to standard hematoxylin and eosin (H&E) histochemical staining as well as 24-plex cyclic immunofluorescence (CyCIF) encompassing >10^7 cells. The sample was systematically analyzed via histopathologic examination by a pathologist, single cell segmentation and classification of cell-states by marker expression, and machine learning-based approaches to predict cell-state transitions and spatial relationships at both single cell and tissue-level resolutions. These analyses revealed spatially-organized molecular and morphologic transitions and spatial interactions occurring between diverse cell populations in distinct tumor microenvironments. Furthermore, they demonstrate the advantages of whole-slide 2D and 3D imaging to understand tumor biology at different levels of tissue organization.