Piconewton (pN) forces regulate many biological processes, including cell migration, immune function, and hemostasis. Molecular tension probes (MTPs) quantify forces cells transmit to their extracellular environment via force-induced unquenching of a fluorophore-quencher pair flanking a force-extensible polymer, transducing pN forces into fluorescence and enabling microscopy measurement of receptor force distribution and magnitude. Despite their unique capabilities, MTPs are challenging to use and synthesize, limiting widespread adoption. Furthermore, only forces transmitted through MTPs are reported, complicating their use in native extracellular matrix. We hypothesized that imaging of cellular morphology and the spatial distribution of force-transducing proteins provides sufficient information to predict pN receptor tension maps. We designed deep learning approaches based on U-Net architectures to predict spatial maps of pN receptor tension from fluorescence images of GFP-tagged vinculin, a key force transducing protein. We named our model tension deep-learning (TensionDL). To train the TensionDL model, we curated a dataset with ∼1,700 images of fibroblasts stably expressing GFP-vinculin. We conducted extensive experimental validation to calibrate predictions and to ensure generalization to new image inputs. TensionDL accurately predicts the distribution of pN receptor forces and is compatible with both immunostaining and live-cell measurements of force-transducing proteins. We demonstrate the accuracy of TensionDL at the subcellular and cellular scale, and have successfully employed TensionDL to predict fibroblast, epithelial cell, and lymphocyte tension images, suggesting our technique works across different biological systems. TensionDL generalizes predicting forces transmitted by cells to soft, hydrogel substrates and successfully inferred tension distribution on fibronectin coated surfaces in which forces are not transduced through MTPs. In conclusion, our results suggest cell morphology and the spatial distribution of mechanically active proteins encode information about receptor mechanics, paving the way to inferring pN receptor forces in heterogeneous environments and, eventually, in natural tissues.
Kansal et al. (Sun,) studied this question.