Predicting nuclear shape from cell shape in mammalian cells

The cell is an internally tensed blob of matter, with an elaborate dynamic polymeric network, the cellular cytoskeleton, that spans the cellular interior. The cytoskeleton is under tension due to the activity of myosin motors, and experimental evidence indicates that these cytoskeletal forces play a role in controlling the shape of the nucleus, that in turn impacts gene expression and cellular phenotypes. Cytoskeletal forces also determine the shape of the cell. We speculated that if the cytoskeleton mechanically couples cell shape and nuclear shape, the latter should be predictable from the former. We first analyzed a publicly available dataset of tens of thousands of cells and nuclei using Deep Neural Nets and a custom Convolutional Neural Net. Our results show that many features of nuclear shape can be predicted with a high accuracy just using the shape of the cell membrane. We next performed imaging experiments on two cells lines, HeLa cells and NIH 3T3 fibroblasts, where we perturbed cytoskeletal prestress using pharmacological agents. Machine-Learning models continued to perform well on both control and the perturbed cells, but we observed interesting differences, which provide insights into the cell-nucleus mechanical coupling and have implications for coarse-grained mechanical models of the nucleus.