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 less is known about the relevant collective variables or the relationships between them. Such problems arise in embryogenesis, where genetic control, force-generating motor proteins, and the system geometry drive dramatic rearrangements of tissue layers. Here, we show that deep neural networks can learn to predict such tissue flows in developing Drosophila embryos using cytoskeletal protein distributions. Our machine learning models are capable of determining the instantaneous tissue velocity as well as how the tissue will deform for several minutes into the future. Using a geometric data-driven approach, we can identify spatial regions whose protein distributions are crucial for determining embryo behavior. Our study incorporates neural networks as an integral part of a framework for constructing predictive phenomenological models in biology and biophysics.