Computational methods for multiplex image analysis of the tumor microenvironment

PROJECT SUMMARY Spatial biology techniques, such as multiplex immunofluorescence (mIF) imaging, enable detailed molecular profiling of tissue microenvironments while preserving spatial context. These techniques hold tremendous potential in basic research and personalized medicine, but the complexity of the resulting data necessitates carefully-designed computational methods to fully leverage this potential. Such approaches are especially needed for studying the tumor microenvironment, where both the cellular composition and spatial structure may play a critical role in patient outcomes. Our project will combine a unique mIF dataset with biologically-informed computational methods to further our basic understanding of the tumor microenvironment and identify patterns that can ultimately inform new treatment strategies. The dataset is derived from a clinically validated mIF assay that was performed prospectively across a real-world cohort of 2,032 patients encompassing over 20 cancer types. We will develop several complementary computational strategies tailored to multiplex images, ranging from spatial statistical methods to uncover tumor phenotypes in an unsupervised fashion, to deep learning models trained to directly predict patient outcomes. Importantly, the methods are designed with interpretability in mind, offering opportunities to identify spatial biomarkers associated with tumor, genomic, and clinical factors. Towards improved scalability, we will also develop deep learning approaches to predict mIF-based phenotypes from routine histopathology slides. The strategies developed in this project will be broadly applicable across spatial biology applications, and we will make these tools freely available to support other researchers.