Date of Award

May 2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Biomedical Sciences

First Advisor

Motoki Takaku

Abstract

Conventional cancer diagnosis and monitoring rely on imaging and tissue biopsies. Imaging is constrained by resolution, while biopsies are invasive, limited in capturing tumor heterogeneity, and challenging to perform repeatedly over time. Liquid biopsy has emerged as a minimally invasive alternative that enables real-time assessment of tumor-derived material in circulation. Among its components, cell-free DNA (cfDNA) has gained significant attention as a promising biomarker for cancer detection and disease monitoring. However, most current cfDNA approaches rely on mutation- and methylation-based signals, which, while informative, can be limited by low tumor fraction, biological heterogeneity, and reduced sensitivity in early-stage disease. In contrast, chromatin-informed fragmentation features remain comparatively underutilized despite their potential to provide complementary biological insight. Increasing evidence suggests that cfDNA fragmentation is not random but reflects the underlying chromatin organization of the tissues of origin. In particular, nucleosome positioning and chromatin accessibility shape fragmentation patterns, resulting in preferential enrichment of cfDNA fragments at specific genomic regions. Building on this concept, this dissertation investigates whether cancer-associated cfDNA signals are preferentially enriched at open regulatory chromatin regions. By anchoring cfDNA analysis to biologically defined reference maps of open chromatin, this work reframes cfDNA as a structured epigenomic signal, providing a biologically grounded framework for more accurate and interpretable detection of cancer-associated changes. To test this framework, this study integrates whole-genome cfDNA sequencing with computational analysis across multiple experimental settings, including cancer versus healthy comparisons, cross-cancer validation, and longitudinal treatment response profiling. cfDNA fragmentation signals were quantified over predefined reference open chromatin genomic loci derived from tumor and immune cell types, enabling analysis within biologically relevant regulatory contexts. Initial analyses employed count-based statistical methods using negative binomial models to identify differentially enriched regions between cancer and healthy samples. While these univariate approaches are effective for detecting strong individual signals, they are limited in capturing coordinated patterns across multiple genomic regions and are sensitive to noise and multiple testing constraints, particularly in low tumor fraction settings. To address these limitations, differential analysis and downstream pathway annotation were complemented by XGBoost, a machine learning model that leverages chromatin-guided features to capture complex, non-linear relationships across genomic regions. Importantly, because the model is trained on predefined open chromatin loci, its predictions remain biologically interpretable, allowing key genomic regions to be mapped to their nearest genes and associated regulatory elements. This integrative strategy emphasizes that the selection of reference open chromatin loci defines the analytical feature space, enabling the extraction of meaningful, robust, and interpretable cfDNA signals that can distinguish cancer states and reveal biologically relevant patterns across datasets. Across both in vitro and patient-derived samples, cfDNA fragmentation patterns exhibited canonical nucleosome-associated features, including dominant mono-nucleosomal peaks and periodic cleavage signatures, confirming that circulating DNA retains structured chromatin-derived information. When analyzed within reference open chromatin loci, cfDNA signals were significantly enriched at promoters and enhancers, with differential regions between cancer and healthy samples mapping to biologically relevant regulatory elements. Integration of tumor-derived and immune-associated open chromatin regions enabled the identification of a refined set of discriminatory loci that robustly separated cancer from healthy samples. These chromatin-informed features were then used to train XGBoost models, which achieved strong classification performance in independent publicly available breast cancer and healthy donor cfDNA datasets. This framework was further extended to pancreatic and ovarian cancer using disease-relevant open chromatin reference regions. Beyond cross-sectional classification, chromatin-informed cfDNA analysis in metastatic breast cancer further revealed dynamic changes associated with therapeutic intervention, where pre-treatment cfDNA profiles distinguished long-term and short-term responders to CDK4/6 inhibition. Longitudinal analysis showed that both tumor-associated and immune-associated chromatin signals evolve during therapy, reflecting underlying regulatory and epigenetic remodeling processes. Together, these findings demonstrate that reference-guided analysis of open chromatin loci enables robust, interpretable classification and provides a unified framework for cancer detection, disease monitoring, and therapeutic response prediction.

Download

Share

COinS