Towards a Framework For Self-Supervised Graph-Based Machine Learning

Data represented in non-euclidean space in the form of graphs provide invaluable insight into objects and their relationships with neighbours. We start with an assay of unfused cells which forms our initial graph--the nuclei are our nodes. Subsequent edges are formed by implementing the delaunay triangulation algorithm on the assay.

Graph Convolutional Networks are used to analyze the topological properties of this graph, enabling the prediction of spatial patterns such as syncytial-dominated or unfused-dominated regions. By leveraging self-supervised learning on small subgraphs--given that our nodes carry no labels--and higher-order neighbor aggregation, these networks aim to uncover consistent structural features, offering a deeper understanding of spatial heterogeneity.

My work will continue to explore the integration of persistent homology and advanced GCN architectures to enhance the detection of spatial dynamics in biological systems. Future efforts will focus on validating these models with Agent-Based Simulations (you can find a simple simulation here) and expanding to diverse datasets, aiming to contribute meaningful insights to graph-based machine learning.