Discovering dynamical parameters of synthetic multicellular systems from image sequences

Cells in multicellular systems have complex shapes, yet most techniques for predicting biological dynamics from image sequences discard shape information by reducing cells to point particles. Here, we present a framework that moves beyond point representations by learning dynamic, shape-aware representations of synthetic cells in a multicellular collective directly from 2D time-series images. Specifically, our GraphDINO framework combines a frozen DINOv3 backbone for shape representation with a graph neural network trained through next-frame prediction. GraphDINO learns a per-cell latent representation that captures underlying cellular properties such as growth rate, membrane tension, and cell-cell adhesion without parameter supervision. The quality of this recovery depends on the encoder, with DINOv3 features providing a better shape representation than classical shape encoders. Together, these results demonstrate that combining foundation model features with graph-based interaction rules allows the discovery of key dynamical parameters that govern cellular behavior, directly from image sequences.