Visualization and Interpretability for Multiplexed Spatial Biology

Abstract

Spatial biology investigates how cells organize into tissues and how that organization shapes function, disease progression, and response to therapy. Multiplexed tissue imaging makes these relationships visible by measuring dozens of markers in situ across whole-slide specimens, yielding gigapixel images with millions of cells. Analyzing these data at cohort scale is challenging: images are high dimensional and heterogeneous, and current workflows often separate computational results from the tissue context that gives them meaning. Extracting insight at this scale requires visualization, interpretability, and analysis methods that keep multivariate measurements and AI model outputs grounded in tissue context and enable reproducible cohort-scale comparisons. This dissertation develops methods that make multiplexed spatial imaging data more interpretable, reproducible, and scalable. We introduce psudo, a method and tool for principled pseudocoloring and palette assignment that optimizes color choices under perceptual and spatial criteria to produce faithful, task-appropriate composites. We present Visinity, a visual analytics system for cohort-scale neighborhood analysis that quantifies cell–cell interactions and supports exploratory and confirmatory analysis of recurrent interaction patterns. We contribute SEAL, an embedding interpretability approach that links 2D projections to their spatial images and underlying feature values, enabling users to interpret clusters and trace selections back to tissue context. Across studies with domain experts on human tonsil, mouse lung, and colorectal cancer data, these systems accelerated hypothesis generation and testing and improved communication of results. Together, they shorten the path from measurement to insight and expand the scope of questions that spatial biology can address.