Unsupervised cellular phenotypic hierarchy enables spatial intratumor heterogeneity characterization, recurrence-associated microdomains discovery, and harnesses network biology from hyperplexed in-situ fluorescence images of colorectal carcinoma

Abstract

LEAPH is an unsupervised machine learning algorithm for characterizing in situ phenotypic heterogeneity in tissue samples. LEAPH builds a phenotypic hierarchy of cell types, cell states and their spatial configurations. The recursive modeling steps involve determining cell types with low-ranked mixtures of factor analyzers and optimizing cell states with spatial regularization. We applied LEAPH to hyperplexed (51 biomarkers) immunofluorescence images of colorectal carcinoma primary tumors (N=213). LEAPH, combined with pointwise mutual information (PMI), enables the discovery of phenotypically distinct microdomains, composed of spatially configured computational phenotypes. LEAPH identified a subset of microdomains visualized as the spatial configuration of recurrence-specific signaling networks whose intracellular and intercellular interactions support cancer stem cell maintenance and immunosuppression in the evolving tumor microenvironment. The LEAPH framework, when combined with microdomain discovery and microdomain-specific network biology, has the potential to provide insights into pathophysiological mechanisms, identify novel drug targets and inform therapeutic strategies for individual patients.