Data-driven modeling of the actin cytoskeleton

The actin cytoskeleton is an active, adaptive structure that regulates its own mechanical properties via mechanochemical feedback loops. Aided by a myriad of mechanosensitive proteins, actin filaments organize into stress fibers anchored by focal adhesions when the cell is subject to external stress. While several mechanosensitive proteins have been identified, a comprehensive picture of how they interact with each other and contribute to the cytoskeleton’s dynamics is lacking. We leverage the rich in vivo spatial variation of actin and actin-binding proteins to build models of how these proteins interact. Using U-Net convolutional neural networks trained on images of immunostained cytoskeletal proteins, we quantify the extent to which these proteins can predict each others’ spatial distributions as well as the distribution of traction forces. This framework provides a quantitative method for comparing the force-sensitivity of LIM-domain actin-binding proteins such as zyxin, and is a first step towards building data-driven mechanochemical models of the cell.