Novel application for graph neural networks for cell surface structure detection

Session

Poster Session

Authors

Edward Offord (1)

Affiliations

1. University of Warwick

Keywords

Graph neural networks, cell membrane, lattice light sheet

Abstract text

Summary

Recent advancements in lattice lightsheet microscopy allow for the study of subcellular membrane structure analysis. Precise cell segmentation methods utilize a random walker technique to reconstruct a closed 2D surface of the cell membrane. Cell surfaces are represented as triangulated mesh data. These surfaces can leverage the automated methods of graph neural networks (GNN) to identify regions of interest on the surface, such as but not limited to filopodia, blebs or macropinocytic cups. Points on the mesh are predicted with a binary classification of being inside or outside the desired feature boundary. In this example we focus on the formation of macropinocytic cups formed in the process of macropinocytosis. Results from the GNN show an effective automated identification of regions of macropinocytic cups that outperforms automated mathematical models with equivalent performance to semi-manual and manual labelling methods.

Introduction

Our method focuses primarily on process macropinocytosis in Dictyostelium cells but equal can be applied to other cell surface structures. Cups are formed in the process of Macropinocytosis a method of non-specific extracellular fluid uptake.  Macropinocytosis is used by cancer cells for feeding and is a  method viruses and bacteria can hijack to enter cells, but despite its medical relevance it has been poorly studied so far, owing to its complex 3D dynamics making analysis difficult.

A curvature-enhanced random walker segments individual cells and forms closed 2D surfaces of individual cells as a triangulated mesh.  This mesh can then be converted into a graph recording each vertex of the mesh as a vertex of the graph and edges recording the triangular faces. Each vertex of the graph has an associated feature vector recording the x,y,z coordinates, mean curvature, and fluorescent markers at that point.

Figure 1 shows the example input to the graph neural network.

 

Methods

Labels for the dataset are generated using a method of manual labeling generated using a virtual reality labeling tool Lutton et al [1].  In addition, a semi-manual method based on manually selected cup centres and mathematical model function as silver quality labels.  The labels are vertex-wise binary classification of being inside a cup boundary or outside the boundary. Using these labels and the triangulated mesh a graph can be constructed that attributes a feature vector and label to each vertex and edges based on the faces relations. Graphs generated with this method provide the basis for the application of graph neural networks.

Out put of the trained graph neural network is a probability based on the likelihood of a vertex existing inside or outside a macropinosome (or the structure of interest). From these predictions it is possible to extract regions of interest for further isolated analysis.

Results

Using the described method to extract macropinocytic cups it becomes possible to track these cups over time, analysis fluorescent marker localisation, and automate cup counting. Figure 2 shows an example output of the graph neural network.

Conclusion

The work presented provides the basis for a novel application of graph neural networks to cell surface data that aims to automate the tedious process of extracting interesting features on the cell surface. Future work will focus on applying the technique to a variety of data including filopodia, llamepodia, and multi-cellular development. Additionally the inclusion of time as a feature can allow for the automation of life-cycle classification, for example identifying the part of the life-cycle the macropinosome is at.

More information on the methods described Offord et al [2].

References

[1] Adam Platt, Judith Lutton, Edward Offord, and Till Bretschneider, “Micellanngelo: annotate microscopy time series of complex cell surfaces with 3d virtual reality,” Bioinformatics, vol. 39, no. 1, pp.btad013, 2023

[2]  Edward Offord, Judith. Lutton, and Till Bretschneider, “Cell membrane feature detection using graph neural networks,” in 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), 2023, pp. 1–5.