Abstract
Lung cancer remains a leading cause of cancer-related mortality worldwide, with survival outcomes strongly dependent on the stage of diagnosis, tumour biology, and access to effective treatment. Since diagnosing lung cancer at an early stage is associated with improved clinical outcomes, identifying reliable biomarkers for early detection remains a clinical priority. Cell-free DNA (cfDNA) methylation analysis has emerged as a promising minimally invasive approach for lung cancer detection, particularly in settings where tissue biopsies or Low-dose computed tomography (LDCT) scans are unsuitable or inconclusive. However, the translational application of cfDNA methylation is complicated by the low abundance of cfDNA. This thesis addresses these challenges through five interconnected themes and aims to develop an early diagnostic method for lung cancer.
The first theme systematically evaluated methods for generating reproducible cfDNA methylation libraries. To achieve this, cfDNA methylation libraries were prepared using reduced representation bisulfite sequencing (RRBS), enzymatic methylation sequencing (EM-Seq), and cell-free reduced representation bisulfite sequencing (cfRRBS). These methods were benchmarked against conventional genomic DNA methylation libraries and external datasets generated using cell-free methylation sequencing (cfMethyl-Seq). In addition to offering a cost-effective alternative, the analyses demonstrated that cfRRBS represents the cfDNA analogue of conventional RRBS on genomic DNA, outperforming EM-Seq on key analytical metrics while showing comparable performance to cfMethyl-Seq. Therefore, cfRRBS was used to prepare the cfDNA methylation libraries in this work.
The second theme focused on the bioinformatic interpretation of cfDNA methylation data. Since differential methylation analysis is highly dependent on statistical modelling strategies, commonly used tools including DSS, DiffMethylTools, MethylKit, BSseq, and DMAP were systematically evaluated using the leave-one-method-out (LOMO) consensus and a priori biological plausibility approaches. DiffMethylTools demonstrated the strongest consensus-based performance, whereas DMAP identified the most biologically plausible methylation signatures.
The third theme characterised the clinical cohort and pre-sequencing cfDNA features. A total of 248 participants were recruited across multiple sites in Australia and New Zealand. cfDNA quantity, quality, and fragmentation patterns were evaluated in relation to tumour size and disease stage. These analyses demonstrated that cfDNA shedding increased with tumour size and disease severity, with late-stage disease associated with both higher cfDNA concentrations and shorter cfDNA fragments compared to early-stage disease. Although these fragmentation and concentration metrics showed biological relevance, they were insufficient as standalone diagnostic biomarkers for lung cancer.
The fourth theme aimed to identify differentially methylated regions (DMRs) between the control and case groups. To preserve biological relevance, differential methylation analyses were performed using clinically relevant non-malignant controls. The analyses included control versus all-stage cancer, control versus early-stage lung cancer, control versus late-stage disease, and early-stage versus late-stage lung cancer. Gene ontology and regulatory enrichment analyses confirmed the biological relevance of the identified DMRs, while validation was performed using two external datasets.
Finally, the fifth theme developed a read-level methylation framework, CancerFinder, for lung cancer classification using the DMRs identified in this work. Model performance was evaluated using standard classification metrics and benchmarked against open-source tools, including CancerDetector and ichorCNA. CancerFinder demonstrated strong discrimination in the internally validated cohorts, with AUCs of 0.909 – 0.940; however, its performance decreased substantially in external validation, where CancerDetector demonstrated superior discrimination.
Overall, this thesis presents a technical optimisation of the standard cfRRBS workflow, incorporates a clinically relevant non-malignant category within the control group, and applies a sequencing read-level modelling approach for lung cancer prediction. The findings presented in this work advance the translational potential of cfDNA methylation assays for early detection of lung cancer. However, clinical diagnostic utility has not yet been fully demonstrated and will require prospective validation in larger, independently recruited and clinically representative cohorts.