Background
Multiplexed immunofluorescence (mIF) methods image dozens of molecules at subcellular resolution and whole-slide scale, allowing detailed characterization of tumor microenvironments. Although cells exhibit differential subcellular localization of proteins and other molecules depending on their functional states,1–3 these morphological features are difficult to quantify. They are discarded in standard analysis workflows. In this work, we develop a metric defining surface protein polarity from mIF images and apply it to enhance immune cell phenotyping. We define six cell subtypes relating to morphology and measure their relation to patient outcomes.
Methods
Our polarity quantification is described in figure 1. Briefly, we apply a polar transformation to cell-centered image patches, denoise, and take the area-under-the-signal-curve, resulting in a polarity score (range 0–1). The score is thresholded to produce a classification into polar and non-polar cells.
We train two models: a 3-layer multilayer perceptron (MLP) neural network that accepts the percent composition of cell types per sample and predicts binary 60-month survival; and a graph-based neural network4 that accepts 3-hop cell neighborhoods and predicts sample-level survival status.
Results
We verify that polarity events (figure 1) are correlated with positive patient outcomes, even after controlling for technical artifacts. Introducing polarity-aware cell subtypes improves model performance in predicting survival status across three patient cohorts (table 1). Given that polar expression can indicate antigen engagement during contact with tumor cells,5 we confirm that polar cells are more likely than uniform cells to be adjacent to tumor cells; conversely, cells with tumor cells as direct neighbors are more likely to be polar. Tumor cells are significantly more likely to neighbor polar than non-polar immune cells (figure 2).
Finally, keeping the number of polar cells constant, we permute the spatial arrangement of polar cells. This significantly influences survival predictions (figure 3), suggesting that polar cell contacts, not simply their presence in a sample, are important.
Conclusions
We describe a robust, interpretable subcellular morphology metric that reflects macrobiological states. Although labeling polarity events as immune synapses is premature, they represent biologically relevant signals in tumor microenvironments.
References
Weber MS, Prod’homme T, Patarroyo JC, Molnarfi N, Karnezis T, Lehmann-Horn K, et al. B-cell activation influences T-cell polarization and outcome of anti-CD20 B-cell depletion in central nervous system autoimmunity. Ann Neurol. 2010 Sep;68(3):369–83.
Garcia E, Ismail S. Spatiotemporal regulation of signaling: focus on T cell activation and the immunological synapse. Int J Mol Sci. 2020 May 6;21(9).
Negulescu PA, Krasieva TB, Khan A, Kerschbaum HH, Cahalan MD. Polarity of T cell shape, motility, and sensitivity to antigen. Immunity. 1996 May;4(5):421–30.
Wu Z, Trevino AE, Wu E, Swanson K, Kim HJ, D’Angio HB, et al. SPACE-GM: geometric deep learning of disease-associated microenvironments from multiplex spatial protein profiles. BioRxiv. 2022 May 13;
Juzans M, Cuche C, Di Bartolo V, Alcover A. Imaging polarized granule release at the cytotoxic T cell immunological synapse using TIRF microscopy: Control by polarity regulators. Methods Cell Biol. 2023;173:1–13.
Abstract 1314 Figure 1
Workflow of computing polarity and Samples of polar and uniform cells
Abstract 1314 Figure 2
Neighborhood composition of polar vs uniform T and B cells. Tumors are significantly more present next to polar cells vs uniform cells. Only neighbor cell types with significant differences between polar and uniform are shown
Abstract 1314 Figure 3
Permutation experiments: Sample 7-hop neighborhood permutations for CD8 T cells. Left is for polar CD8+ T cells, and right is for uniform CD8+ T cells
Abstract 1314 Table 1
Prediction performance on patient survival status increases when introducing the additional polarity-derived cell types. This effect is observed across multiple datasets and disease types