Graph Neural Networks for Physical Models of Collective Cell Migration

Abstract

In this thesis, a graph neural network generates likely distributions of the set of points corresponding to the positions of cells after a discrete time step from a particular initial disposition. The predictions of this probabilistic program will be sampled to be used over and over again to simulate the trajectories of each cell. From this generated set of trajectories, summary statistics that reflect the expected behavior will be compared against the same statistics computed on 2D synthetic data generated from a cell migration simulator and then real, computer annotated, data coming from roaming neural crest cells in a dish. Finally, the mechanics of the model will be analyzed in order to collect an understanding of its decision processes, which will be compared to the known mechanics of the simulator. It will be shown that it is possible to train a small and scalable model to produce accurate trajectories for most of the scenarios studied and that, with only a few real, computer annotated, samples, this model still offered interesting inference capabilities. The explainability of graph neural networks and attention layers will be leveraged to offer some insight on the decision processes. The use of such models can contribute to improving automatic data annotation and provide alternate angles to study the mechanisms of cell migration.