Learning Mechanisms for Collective Multicellular Behavior with Differentiable Molecular Dynamics Simulations

Living cells display a remarkable ability to self organize into increasingly complex structures - from "symmetry breaking" of identical cells in the embryo to the formation of organoids in vitro. During development, cells can undergo a complex sequence of intercellular interactions and movements to create spatiotemporal organization. However, uncovering the developmental programs that orchestrate this multicellular behavior has proven to be a challenge. In this work, we leverage advances in machine learning technologies to optimize over physics and biology based molecular dynamics (MD) simulations of individual cells, to learn mechanisms that can drive collective behavior. We apply this framework to recover cell-based rules to drive (i) homogeneous tissue growth, (ii) homeostasis between cell types and (iiiI) elongation and sorting of a symmetric cluster of cells. Our work opens new avenues for designing cellular interactions to program complex multicellular behavior, as well as to learn mechanisms from experimental data of developmental trajectories.