A flexible systems analysis pipeline for elucidating spatial relationships in the tumor microenvironment linked with cellular phenotypes and patient-level features

Gabriel F Hanson¹Kate A Goundry¹Remziye E Wessel¹Michael G Brown^2,3,4Timothy NJ Bullock^2,5,6Sepideh Dolatshahi^1,2*

• ¹University of Virginia Department of Biomedical Engineering, Charlottesville, United States
• ²University of Virginia Beirne B Carter Center for Immunology Research, Charlottesville, United States
• ³University of Virginia Department of Medicine, Charlottesville, United States
• ⁴University of Virginia Department of Microbiology Immunology and Cancer Biology, Charlottesville, United States
• ⁵University of Virginia, Charlottesville, United States
• ⁶University of Virginia, Department of Pathology, Charlottesville, VA, United States

Introduction: Quantitative investigation of how the spatial organization of cells within the tumor microenvironment associates with disease progression, patient outcomes, and that cell’s phenotypic state remains a key challenge in cancer biology. High-dimensional multiplexed imaging offers an opportunity to explore these relationships at single-cell resolution. Methods: We developed a computational pipeline to quantify and analyze the neighborhood profiles of individual cells in multiplexed immunofluorescence images. The pipeline characterizes spatial co-localization patterns within the tumor microenvironment and applies interpretable supervised machine learning models, specifically orthogonal partial least squares analysis (OPLS), to identify spatial relationships predictive of cell states and clinical phenotypes. Results: We applied this framework to a previously published non-small cell lung cancer (NSCLC) cohort across four applications. At the cellular level, we identified neighborhood features associated with lymphocyte activation states. At the tumor-immune interface, we demonstrated that the immune cell composition surrounding major histocompatibility complex class I-expressing (MHC I+) tumor cells could distinguish adenocarcinoma from squamous cell carcinoma. At the patient level, spatial features predicted tumor grade. Discussion: By integrating cell-segmented imaging data with interpretable modeling, our pipeline reveals key spatial determinants of tumor biology. These findings generate testable mechanistic hypotheses about intercellular interactions and support the development of spatially informed prognostic and therapeutic strategies.