Automatic Tumour Typing based on Patterns of Somatic Passenger Mutations

In cancer, a tumour’s cell of origin is the strongest determinant of its clinical behaviour. While cell of origin is typically clear at the time of diagnosis, 3-5% of cancer patients present with a metastatic tumour and no obvious corresponding primary tumour. Despite advances in molecular testing, imaging, and pathology, the primary tumour site cannot be inferred in the majority of these cases. Recent large- scale analysis of cancer genomes has uncovered strong associations between cancer type and somatic mutations, prompting the use of somatic mutations as a tool for identifying cancer type. While existing approaches have attempted to use cancer-associated mutations, which may be more common in specific cancer types to infer the primary tumour type from the metastatic tissue, these methods have had only limited success. A more promising alternative is to use the association between patterns of somatic passenger mutations and cancer type, by exploiting the relationships between both regional mutation density and cancer type, and mutational processes and cancer type. Somatic point mutations accu- mulate in regions of closed chromatin, and so mutation density provides information about chromatin state, which in turn offers hints about the underlying cell type. As some mutational processes are highly cell-type specific, mutational processes also provide clues about cancer type. In this thesis, I describe a number of deep learning systems for automatic tumour typing based on patterns of somatic passenger mutations. I then address challenges for translating the classifier into clinical scenarios through the use of multiple algorithmic improvements. First, I make use of modern advancements in deep learning to extend the classifier to accurately discriminate between 29 cancer types. I then use a number of sta- tistical methods for assessing the uncertainty in the model’s predictions, and for improving uncertainty quantification. Finally, I make use of information theoretic metrics to use the model’s predictive uncer- tainty to automatically detect cancer samples that come from rare cancer types that the model was not trained to classify. These studies demonstrate the utility of passenger mutations as a tool for identifying cancer type, and address challenges for translating the deep learning classifier into clinical settings.

Department dc:contributor.department

Molecular Genetics

Year dc:date.issued

2021

Author dc:creator

• Atwal, Gurnit

Advisor dc:contributor.advisor

• Morris, Quaid D

• Cancer diagnosis
• Chromatin
• Deep learning
• Somatic mutations
• Uncertainty estimation

Handle dc:identifier.uri

http://hdl.handle.net/1807/123114

OAI identifier oai:identifier

oai:utoronto.scholaris.ca:1807/123114