Biologically Informed Modelling of the Tumour Microenvironment: Hyperbolic Representation Learning and Hierarchical Association Rules

Understanding spatial cell organisation in the tumour microenvironment is critical for cancer research, as it provides key insights into disease mechanisms. In this ongoing work, we explore two approaches that incorporate biological knowledge while analysing the tumour microenvironment. Firstly, we embed single-cell instances into a Lorentzian latent space using a fully hyperbolic variational autoencoder, which we have previously shown to retain significantly more biologically relevant information than Euclidean alternatives. Secondly, we adopt taxonomy-aware spatial association rule mining for quantifying cell interaction behaviour beyond pairwise interactions.