Machine learning Drosophila embryogenesis

Hydrodynamic theories use symmetries and conservation laws to build effective descriptions of many-body systems. This approach can break down in living and active matter, where it is difficult to identify relevant collective variables and the lack of symmetries leaves the functional relationships between variables unconstrained. Such problems arise in embryogenesis, where force-generating motor proteins drive dramatic rearrangements of tissue layers. Here, we show that deep neural networks can learn to map between the time-varying distribution of myosin motors and the cellular flows occurring in developing Drosophila embryos. Our models account for the curved geometry of the embryo and can reconstruct both the cellular flows and myosin configurations with high accuracy at different phases of gastrulation.