T20 - Machine Learning of EMT State Landscapes via Live Imaging of Collective and Individual Migration

Introduction: : Functional cell states and state transitions are shaped by interactions with neighboring cells, driving emergent behaviors such as collective migration and spatial patterning. In particular, the epithelial-mesenchymal transition (EMT) is associated with transitions from adherent tissues towards collective and individual migration during embryonic development, wound healing and tumor progression. These transitions could occur via, partial or hybrid EMT states that retain some cell-cell adhesions and act as “leader cells” to coordinate collective migration. However, profiling these partial EMT states remains challenging due to the heterogeneity and plasticity of migratory cells. Here, we utilize live-cell imaging and machine learning to identify a landscape of EMT states and how heterogeneous cells act in concert to coordinate motility. We train convolutional neural networks with single cell readouts of a fluorescent EMT reporter, morphology and migration to reveal a cell state manifold beyond molecular biomarker expression.

Materials and

Methods: : Mammary epithelial cells (MCF10A) with a genetically encoded EMT fluorescence reporter (Z-cad) were a generous gift from MJ Toneff and JM Rosen (UT Southwestern). These cells express red fluorescent protein (RFP) driven by an E-cadherin (CDH1) promoter, which switches to green fluorescent protein (GFP) regulated by ZEB1 3’ UTR after EMT induction. Cells were further labeled with CellTracker FarRed to display partial states in the absence of RFP or GFP expression. Epithelial cells were treated with TGF-beta1 to induce partial or complete EMT at varying concentrations (1-5 ng/mL for 7 days). These various types of epithelial cells were cultured in standard media on fibronectin-coated, and subsequently imaged using live-cell microscopy (Figure 1A). Cell morphologies were analyzed using Cellpose for the Cytoplasmic channel, while collective migration was assessed by applying the optical flow algorithm to sequential images of cell nuclei stained with Hoechst 33342. Cells were also plated on compliant polyacrylamide substrates coated with collagen I for traction force microscopy.

Results, Conclusions, and Discussions:: We analyzed the expression of EMT fluorescent reporters across different concentrations of TGF-beta1 to examine the reliability of reporter expression in capturing the EMT states. As shown in Figure 1B, TGF-beta1-induced EMT led to significant changes in reporter expression at both mild (1ng/ml) and intense (5ng/ml) concentrations. The subpopulation exhibiting RFP signal decreased markedly upon EMT induction, whereas the GFP-positive population increased. Notably, we found the number of non-colored cells increased significantly in TGF-beta-treated samples, demonstrating the EMT states of cells cannot be binarized as purely epithelial or mesenchymal types. Moreover, although morphology is a well-established hallmark of EMT, cells with the same reporter expression still displayed critical morphological diversity, highlighting the limitations of reporter-based classification alone (Figure 1C). In response, we utilized the machine learning (ML)-based segmentation and extracted 46 morphological features, including basic and moment-based descriptors, and constructed the morphospace after dimension reduction using principal component analysis (PCA). From the most distinct cell groups, confluent epithelial cells and sparse mesenchymal cells, we soft clustered cells into three groups roughly correlating with the size of cells (Figure 1D). By combining the clusters and color expression, we could classify the 12 groups of subpopulations for the EMT states (Figure 1E). Remarkably, a group of cells with intermediate size and E-cadherin expression (RFP) from mild EMT induction (1ng/ml TGF-beta1) showed faster and continuous migration, suggesting a “leader” phenotype maintaining the cell-cell junction and greater motility (Figure 2AB). Accordingly, the mixing of such intermediate EMT cells facilitated the faster and coordinated migration by reducing the jamming transition in the normal epithelial cells.
In conclusion, we find that collective cell migration is shaped by the heterogeneity of subpopulations arising from epithelial-mesenchymal plasticity. Our integrated model, which combines morphological features and molecular expression, identified specific subgroups of cells that drive robust migration within the epithelial layer. This framework naturally captures comprehensive cell-level information related to the collision and spatial networks, giving valuable insights into how heterogeneity influences cell migration across diverse cell arrangements, including the embryo and tumor clusters.

Acknowledgements and/or References (Optional):: We acknowledge funding from NIGMS (R01GM140108) and the Brown University Hibbitt Postdoctoral Fellowship and Undergraduate Teaching and Research Award.