A morphology-based machine learning model for scoring epithelial-mesenchymal plasticity using organelle dynamics.

Re-activation of epithelial-mesenchymal transition (EMT), a key developmental process, contributes to cancer progression and therapy resistance. Modulating EMT could be attractive as a therapeutic strategy, but there is a lack of methods that can quantify EMT states, including hybrid phenotypes. Here, we developed a morphology-based machine learning approach to score EMT based on changes in organelle dynamics. Using the Cell Painting assay and high-throughput microscopy, we trained a histogram gradient boosting classifier to identify stage-specific organelle remodeling during a time course of TGF-β1-induced EMT in mammary epithelial cells. The model achieved robust performance across datasets, capturing EMT kinetics, hybrid states, and reversal by mesenchymal-epithelial transition (MET). Importantly, the method accurately scored EMT in human breast cancer cells and lung cancer cells undergoing hypoxia-induced EMT, demonstrating cross-species, cross-inducer, and cross-cancer applicability. The results establish organelle morphology profiling as a scalable framework for quantifying epithelial-mesenchymal plasticity. The method offers a platform for drug discovery and identifying strategies to overcome EMT-associated resistance.