Deep learning-driven morphological analysis for assessing EMT state and drug sensitivity of single tumor cell

Abstract

Metastasis driven by the epithelial-mesenchymal transition (EMT) in circulating tumor cells (CTCs) is a major challenge in cancer treatment. Current EMT assessment methods rely on invasive detection of protein or genetic markers, lack single-cell resolution, and fail to provide real-time dynamic insights, especially for rare CTCs. Here, we developed a convolutional neural network (CNN)-based deep learning model that quantifies EMT states in single or scarce CTCs through non-invasive, label-free morphological profiling. First, TGF-β-stimulated EMT induction in MCF-7 cells was monitored through quantitative assessment of EMT-related protein expression, identifying key transitional timepoints. Then, five distinct morphological states representing EMT progression were selected via combined morphological observation. Finally, cellular images from these states were processed by the developed convolutional neural network (CNN) model, which performs label-free morphological profiling at single-cell resolution. This approach enables real-time, individualized evaluation of metastatic potential, advancing precision diagnostics and therapeutic strategies for cancer management.