A data-driven modeling pipeline was developed to predict cellular forces from images of cytoskeletal proteins, particularly zyxin. Neural networks were trained to predict traction forces from images of zyxin, revealing that a single focal adhesion (FA) protein can generalize to unseen biological regimes. Two approaches were developed: one constrained by physics and another agnostic, to construct data-driven continuum models of cellular forces. Both approaches revealed that cellular forces are encoded by two distinct length scales. The models demonstrated that zyxin alone can predict traction forces accurately, even in new cell types and under biochemical perturbations. The neural networks identified features of cell adhesion and morphology, showing that zyxin distributions encode information about local forces and whole-cell contractility. A physical bottleneck neural network (PBNN) was introduced to incorporate zyxin into continuum mechanical models, learning relationships between proteins and physical parameters to enhance model generalizability. A Green’s function neural network (GFNN) was also developed, revealing long-range interactions and an analytical formula for traction forces. These approaches demonstrated that neural networks can be used to build predictive models of cell biology, revealing insights into the mechanics of cells. The study highlights the potential of machine learning to advance understanding of cell mechanics and biophysics.A data-driven modeling pipeline was developed to predict cellular forces from images of cytoskeletal proteins, particularly zyxin. Neural networks were trained to predict traction forces from images of zyxin, revealing that a single focal adhesion (FA) protein can generalize to unseen biological regimes. Two approaches were developed: one constrained by physics and another agnostic, to construct data-driven continuum models of cellular forces. Both approaches revealed that cellular forces are encoded by two distinct length scales. The models demonstrated that zyxin alone can predict traction forces accurately, even in new cell types and under biochemical perturbations. The neural networks identified features of cell adhesion and morphology, showing that zyxin distributions encode information about local forces and whole-cell contractility. A physical bottleneck neural network (PBNN) was introduced to incorporate zyxin into continuum mechanical models, learning relationships between proteins and physical parameters to enhance model generalizability. A Green’s function neural network (GFNN) was also developed, revealing long-range interactions and an analytical formula for traction forces. These approaches demonstrated that neural networks can be used to build predictive models of cell biology, revealing insights into the mechanics of cells. The study highlights the potential of machine learning to advance understanding of cell mechanics and biophysics.