Noninvasive Assessment of the Tumor Using cfDNA

Next-generation high-throughput sequencing, which is increasingly generating vast amounts of genomic data, offers opportunities for a deeper understanding of the multifaceted nature of cancer and, hence, better patient care. However, the inherently complex and heterogeneous nature of cancer and the significant challenges in the generated data demand advanced data-driven frameworks to decode the molecular underpinnings of cancer. Moreover, the undeniable need for non-invasive approaches presents additional technical challenges in this domain. This dissertation proposes novel frameworks addressing three key challenges in computational oncology. The first study of this dissertation develops a data-driven algorithm to identify stemness signatures in metastatic castration-resistant prostate cancer (mCRPC) plasma cfDNA methylation data utilizing the overall inverse relationship between methylation and gene expression. The second study proposes an efficient, robust, scalable multi-phenotype differentially methylated region (DMR) detection approach. Additionally, this study systematically evaluates deep learning architectures for genomic translation in the context of inferring methylation from urinary cfDNA whole-genome sequencing data. Subsequently, the inferred methylation data is utilized in a cancer classification model to distinguish bladder cancer from healthy samples. Finally, in the third study, a multi-cancer classification framework for genitourinary (GU) cancers based on urine samples is proposed. This framework non-invasively classifies bladder, prostate, and kidney cancers along with healthy samples. The robustness of the framework at varying tumor fractions is validated through in silico simulations. Incorporating sophisticated algorithmic, deep learning, and machine learning approaches, this dissertation proposes frameworks to address three key challenges in advancing non-invasive cancer care. Demonstrating strong performance through extensive experiments, it is intended to serve as an important step toward overcoming current challenges in the development and application of computational methods for cancer management.