Cancer develops over decades before becoming clinically detectable. Understanding this pre-cancerous evolution is crucial for improving the detection, prevention, and treatment of cancer. Previous research has predominantly centred on the somatic and inherited variants which drive cancer. However, altered methylation is also a hallmark of cancer development and is a promising tissue-specific biomarker for early detection. It remains unknown how early in tumour development methylation changes occur, how these early changes progress as the tumour develops, and how methylation dynamics are shaped by the acquisition of somatic driver mutations. Acute Myeloid Leukaemia (AML) is a good model system with which to address these questions. As well as the ease of sampling blood, mutations in epigenetic modifiers are some of the earliest and most common driver events in AML. In this work, we used peripheral blood samples from a longitudinal cohort of 50 pre-AML donors and 50 age-matched controls (>500 samples overall) to understand the dynamics of methylation change through time in the decade preceding diagnosis.

To ensure sensitivity to rare AML-specific molecules and obtain precise and unbiased estimates of methylation frequency over time, a highly sensitive targeted methylation sequencing strategy was needed. To this end, the first aim of this work was to identify differentially methylated regions (DMRs) in AML. In Chapter 2, we developed a statistical method for identifying DMRs from whole genome bisulfite sequencing data and applied this to existing datasets of purified AML blasts and healthy hematopoietic stem cells (HSCs). We identified ~ 6000 high-confidence genomic regions that were differentially methylated in AML. Both hypomethylated and hypermethylated DMRs were strongly enriched in functional regions of the genome. However, the patterns of methylation change for hypermethylated regions were far more AML-specific than for hypomethylated regions. In Chapter 3, we showed that hypermethylated DMRs were able to distinguish pre-AML samples from healthy controls up to 2 years prior to diagnosis. However, we found that this pre-AML signal varied considerably among individuals. We also carried out DMR discovery using the pre-AML samples and matched controls to identify further relevant regions.

In Chapter 4, we developed a custom enrichment panel targeting ~ 5500 regions and a targeted methylation sequencing workflow capable of capturing ~ 1000 unique molecules per site. The panel included the DMRs identified, AML promoters, and clock regions. In Chapter 5, by applying this deep sequencing approach to >250 serial blood samples collected annually in the decades preceding 50 incidental AML diagnoses and age-matched controls, we revealed how DMRs clearly distinguished pre-AML cases from controls up to 14 years prior to a clinical diagnosis. In contrast to the observations in pre-AML cases, the dynamics of methylation in healthy controls were remarkably stable between individuals and through time. By combining longitudinal methylation sequencing with longitudinal somatic variant calls, the abundance of specific somatic driver mutations could be associated with methylation levels at specific CpGs, thus providing a genotype-phenotype map for certain driver mutations. Remarkably, we found that driver mutations in diverse AML-associated genes are associated with a convergent CpG phenotype.

Statistical models for AML risk prediction based solely on methylation signal, namely a binomial model and a logistic regression model, showed sensitivity rates of ~ 30%, at specificity rates of > 97% (Chapter 6). These results highlight the rich levels of information in methylation patterns and show how they could be used for risk stratification very early in cancer development.