Abstract
Epithelial-mesenchymal transition (EMT) occurs heterogeneously among malignant carcinoma cells to promote chemoresistance. Identifying the signaling pathways involved will nominate drug combinations to promote chemoresponse, but cell population-level studies are inherently fraught, and single-cell transcriptomics are limited to indirect ontology-based inferences. To understand EMT heterogeneity at a signaling protein level, we combined iterative indirect immunofluorescence imaging of pancreas cancer cells and tumors and mutual information (MI) modeling. Focusing first on MAP kinase pathways, MI predicted that cell-to-cell variation in ERK activity surprisingly dominated control of EMT heterogeneity in response to diverse growth factors and chemotherapeutics, but that JNK compensated when MEK was inhibited. Population-level models could not capture these experimentally validated MI predictions. The dominant role of ERK was predicted by MI even when analyzing seven potential EMT-regulating signaling nodes. More generally, this work provides an approach for studying highly multivariate signaling/phenotype relationships based on protein measurements in any setting.