Heterogeneous Graph Neural Networks for Analysing Spatio-Temporal Cell Surface Dynamics

Cell surface membrane deformations underpin fundamental biological processes, such as cell division, motility, or uptake of fluid or particles. Fast lightsheet timelapse microscopy enables imaging these dynamic processes in 3D with high spatial and temporal resolution, generating large imaging datasets that can be challenging to analyse, in part due to their size. We introduce heterogeneous graph convolutional neural networks (hGCNs) to analyse dynamic features on triangulated cell surfaces, reducing the dimensionality to 2D + time. Construction of the underlying spatio-temporal hypergraph employs spectral graph correspondence techniques to map surface nodes between time frames. The utility of hGCNs is demonstrated using synthetic data, showing the ability to discriminate between temporally varying features and to identify the growth stage of thin protrusions on the cell surface known as filopodia. We achieve robust performance, with an F1 score of 0.952 when classifying the life cycle of filopodia in simulated lung cancer cells. Our generic approach has potential for adaptation to numerous other dynamic features on cell surfaces.