CoSy seminar with Youssef Wally 15/9
- Date
- to
- Location
- Ångström Laboratory, Å2002
- Type
- Seminar
- Lecturer
- Youssef Wally
- Organiser
- Department of Mathematics
- Contact person
- Jörgen Östensson
Youssef Wally will hold a CoSy seminar
Everyone is welcome to attend and the first 40 people who register will get a free sandwich. Register via the following link by 13/9:
https://doit.medfarm.uu.se/bin/kurt3/kurt/8910615
Title: Biologically Informed Modelling of the Tumour Microenvironment
Abstract:
Modern multiplexed imaging allows us to capture dozens of proteins expressed in tissue, offering an unprecedented understanding of the tumour microenvironment (TME): it lets us study which cells interact with each other and which cells and proteins drive tumour growth; how tumours form, how they spread, and how immune cells attack tumour cells. Extracting this knowledge, however, is far from straightforward. The data is highly complex, and cell classification itself remains a challenging problem, so machine-learning analyses of cell neighbourhoods must rely on noisy cell-type labels. Moreover, these labels are usually treated as an unstructured set of categories, although biologically some cell types are more closely related than others: two specialised immune cells are more related to each other than to a tumour cell. In this seminar, I present two complementary approaches to anchoring this knowledge of how cell labels are taxonomically related in data-driven pattern discovery in cell neighbourhoods. One drawing on non-Euclidean geometry, the other on market basket analysis and logic. Both are evaluated on a breast cancer tissue.
First, since cell types form a taxonomy, we investigate whether these relationships are better represented in a non-Euclidean geometry designed for hierarchical structure. Embedding cells into Euclidean and Lorentzian latent spaces with a fully hyperbolic variational autoencoder and quantifying performance with information-theoretic estimators, we show that hyperbolic embeddings retain significantly more relevant information than their Euclidean counterparts. Second, since cell interactions are inherently directional, a tumour cell signalling to a neighbouring immune cell is not the same event as the reverse, symmetric co-occurrence statistics cannot capture them. We therefore introduce a hierarchy-aware spatial rule-mining framework that discovers asymmetric, beyond pairwise interactions, statistically significant rules of the form "X implies Y" from segmented and classified tissue data, where rules may combine several cell types across different levels of the taxonomy.