Probing the rules of cell coordination in live tissues by interpretable machine learning based on graph neural networks

Abstract

Extracting the rules of cell-to-cell interactions in tissue dynamics is challenging even if high-resolution live microscopy is accessible. In order to seek and compare the different rules enforced in developing and homeostatic tissues, it will be desirable to have a systematic method that outputs the rules of multi-cellular kinetics simply from live images and cell tracks. Here we demonstrate that graph neural network (GNN)-based models can predict cell fate in the mammalian epidermis when spatiotemporal graph constructed from cell tracks and contacts are given as inputs. By extracting the rules learned by GNN, we find neighbor cell fate inductions and inhibitions consistent with previous findings as well as some that have been previously overlooked. This study demonstrates how GNN-based methods can be useful in inferring stochastic dynamics such as multi-cellular kinetics involving proliferation and loss of agents.