Graph neural networks for multicellular dynamics

Cells in a tissue often form a disordered network. It mediates complex intercellular interactions, which strongly affect the migration and division of the tissue cells. Such multicellular dynamics plays a crucial role in many biological processes ranging from wound healing to organogenesis. However, given the complexity of the internal drive of individual cells and their interactions, it is extremely difficult to establish a theoretical model from first principle. Here we propose a generic machine learning approach capable of learning various multicellular dynamics from recorded experimental videos. Instead of requiring the internal states of living cells that are hard to access, our model relies solely on the external geometric information of them (i.e. cell shape, size, and interconnectivity) that are easy to measure. To learn the cell interactions that can be both non-reciprocal and pathway-dependent, we represent a tissue system in terms of both cell and interaction graphs and apply advanced neural networks onto this dual graph representation. Taking epithelial cells as an illustrative example, we show our graph neural network not only captures the stochastic cell motion but also predicts the evolution of cell states in their division cycle. We show this method can be easily extended to forecast the developmental systems (e.g. fly wing cells) and the cell signaling process (e.g. ERK signaling).