Genomic and image-based quantification of intra-tumoural heterogeneity

Thesis

Cancer is a complex dynamic disease which can be explained by an evolutionary process of clonal expansion, genetic diversification, and clonal selection, leading to the presence of clonal heterogeneity in the tumour. Another key aspect of cancer heterogeneity is heterogeneity in the tumour microenvironment. This thesis addresses the challenge of investigating intra-tumour heterogeneity, and proposes statistics and machine learning based methods for exploring variability arising from both genomic and immunological differences, unleashing the potential for improved cancer subtyping.

Genomic sequencing allows us to dive deep into the clonal structure of each individual tumour, permitting the study of specific cell populations arising from distinct mutational processes. On the other hand, the study of whole slide tissue imaging allows us to investigate the tumour microenvironment by exploring the complex interaction between the developing tumour and the immune system.

In this work, I develop a methodology to obtain copy number calls from sequencing data in the more complex case where reads are only available for a set of targeted genes with no matched normal samples and define evolutionary-based clusters in colorectal samples from a large multi-centre consortium. The detected genomic events and resulting clusters are found to be significant for overall survival prediction across the different cohorts.

A second smaller set of colorectal samples is scanned using a novel multiplexed immunofluorescence platform. The analysis of these whole-slide antibody-labeled images involves the development of a graph-based multi-scale deep learning approach to combine single-cell with spatial context information. The model is able to extract features associated with distinct tumour stages. Moreover, the application of interpretability methods allows for the identification of regions with predictive importance. Finally, a third data modality is explored, namely, single-cell fluorescence in situ hybridization imaging, where graph neural networks are again utilised to combine single-cell morphological and genetic information in order to identify differences between non-progressed ductal carcinoma in situ samples and samples progressing to invasive breast cancer, opening up opportunities for early detection.

Authors: García Martín, N

• DOI: 10.5287/ora-o8ygzqbnb
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English