The progression and treatment response of cancer is influenced by the intricate tissue structure where cancer cells are embedded, known as the tumour microenvironment (TME). Recent technical advances allow the acquisition of highly multiplexed biomedical images (HMBI) that generate spatial tissue maps of dozens of proteins capturing the intricacies of the TME.

Combining the multidimensional cell phenotypes acquired with their spatial organization to predict clinically relevant information is a challenging computational task. Inspired by the opportunities that artificial intelligence offers to capture the language of biology, this thesis focuses on using graph representation learning on HMBI to highlight unknown biological patterns in a data-driven manner.

Here I propose MULTIPLAI, a novel framework to predict clinical biomarkers from HMBI data using Graph Neural Networks (GNNs) that integrate both the phenotypic and spatial dimensions of HMBI images while learning the best representation of the whole slide for each task and providing a framework to explore feature attribution. By the time of publication, it was the first application of GNNs to this type of data.

For the first case study of MULTIPLAI, I carry out a proof-of-concept study to predict oestrogen receptor (ER) status, a key clinical variable for breast cancer patients. The results suggest that MULTIPLAI successfully captures TME features with clinical importance.

For the second case study of MULTIPLAI, I explore the relationship between TME patterns and different definitions of chromosomal instability (CIN). The results indicate that a combination of immune cells and markers associated with breast cancer, as well as cellular birth and death processes, play significant roles in different CIN metrics.

These results support the potential of MULTIPLAI for data-driven analysis of HMBI to enhance our comprehension of spatial tumour biology. This will become particularly valuable in a multi-modal research context, where it can shed light on the interconnections between various data sources.