Neural networks for data-driven models of cell mechanics

Mechanical behaviors of cells arise through the mechanochemical interactions of proteins which self-organize into organelles and cytoskeletal structures. However, no systematic strategy exists to identify the relevant collective variables representing protein distributions within the cell and link these to mechanical response at the cellular scale. Here we show how machine learning can link protein distributions to mechanical forces, leading to data-driven physical models without requiring knowledge at the microscale. We train a neural network to predict traction forces in cells from fluorescently-labeled proteins to establish a protein-force relation. From these networks, we extract an effective model of force as a Coulomb-like interaction between localized protein structures within the cell. Next, we construct a data-driven elastic cell model which bypasses microscopic theory and directly links proteins to continuum mechanical parameters. This procedure uncovers length scales of biologically-relevant features of the protein distribution. These models further allow us to perform high-throughput probes of potential biological perturbations to identify promising new experiments.