Single-image inference of clathrin-mediated endocytosis dynamics via deep learning

Clathrin-mediated endocytosis (CME) is a vital cellular process that exhibits spatial and temporal heterogeneity in its dynamics, traditionally studied through labor-intensive time-lapse microscopy and single particle tracking. To overcome the limitations posed by phototoxicity, temporal undersampling, and computational complexity, we introduce a deep learning framework that infers CME dynamics from single fluorescence images. Using a modified U-Net architecture, our model predicts spatial maps of the standard deviation (SD) of clathrin coat growth rates-an established metric of CME activity-directly from static frames of AP2-eGFP-labeled cells. The network was trained on paired image data and SD maps derived from experimentally tracked endocytic events. The model accurately recapitulates dynamic features such as front-rear asymmetry in migrating cells and responses to membrane tension alterations, demonstrating strong agreement with traditional time-lapse-derived metrics. This approach eliminates the need for trajectory reconstruction or prolonged imaging, enabling real-time, non-invasive assessment of endocytic dynamics across diverse biological contexts. Our results highlight the potential of deep learning to extract dynamic biophysical information from static imaging data and establish a scalable methodology for probing CME and related subcellular processes.