Prediction of Piconewton Receptor Tension Images using Deep Learning

Abstract

Piconewton (pN) receptor forces govern many biological processes, but measuring these forces remains challenging. Molecular tension probes (MTPs) provide a sensitive means to measure pN cellular forces via fluorescence microscopy; however, MTPs are challenging to use and only forces transmitted through the probes are reported, complicating their use in heterogenous environments. Here, we present Tension Deep Learning (TensionDL), which leverages convolutional neural networks and image-to-image translation to predict pN receptor tension maps from images of cell morphology and the force-transducing protein vinculin. We validate the accuracy of TensionDL at the subcellular and cellular scales, demonstrate model accuracy across different substrate stiffnesses and cell types, and leverage TensionDL to make semi-quantitative predictions of cell mechanical output. Finally, TensionDL enables long-term mapping of pN receptor tension and infers tension distributions in heterogeneous environments in which some forces are not transduced through MTPs.