Navigating the oncogenomic maze through advanced computational methods

Cancer is a highly heterogeneous disease driven by the accumulation of genomic and transcriptomic alterations and shaped by dynamic interactions between malignant and non-malignant cells within the tumor microenvironment. Advances in high-throughput sequencing technologies have enabled increasingly detailed molecular profiling of tumors, offering unprecedented opportunities for precision oncology. However, the interpretation of these large, noisy, and heterogeneous datasets continues to pose major analytical challenges, especially in solid tumors where both inter- and intra-tumor heterogeneity significantly influence diagnosis, prognosis, and therapeutic response. This thesis presents a collection of computational frameworks that leverage genomic and transcriptomic sequencing data to improve diagnostic and prognostic decision-making in oncology. The methodologies and applications span bulk RNA-sequencing, targeted and whole-exome DNA sequencing, and shallow whole-genome sequencing (sWGS), with a shared focus on enabling robust, clinically actionable insights from routine pathology samples. The first contribution is BLADE (Bayesian Log-normAl DEconvolution), a statistical model designed to accurately infer cellular composition and cell-type-specific gene expression from bulk RNA-seq data, using signatures derived from single-cell RNA-seq. BLADE employs a Bayesian log-normal noise model with scalable variational inference, allowing it to deconvolve complex tumor tissues containing over 20 distinct cell types. Extensive benchmarking demonstrated improved performance over state-of-the-art methods, particularly in recovering cell-type-specific expression, critical for dissecting immune and stromal dynamics that influence treatment response. The framework illustrates how transcriptomic deconvolution can support molecular pathology by enabling clearer interpretation of bulk RNA-seq in clinical contexts. Chapters 3 and 4 address a key diagnostic challenge in oncology: distinguishing metastatic disease from multiple independent primary tumors. Using data from the TRACERx lung cancer study, Chapter 3 shows that current targeted next-generation sequencing (NGS) mutation panels frequently produce inconclusive results when assessing tumor clonality, especially in lung squamous cell carcinoma. Chapter 4 demonstrates that copy-number aberration (CNA) profiling, obtained through sWGS, provides a robust and scalable complementary approach, resolving the majority of ambiguous cases and supporting accurate staging and treatment selection. Together, these studies propose an integrated genomic framework for clonality classification directly applicable in routine pathology workflows. Chapter 5 focuses on clinical prognostication, evaluating whether CNA burden and genomic instability predict distant metastatic recurrence in young women with ER-positive, HER2-negative breast cancer. Using elastic-net regression models across two large patient cohorts, the work reveals that genome-wide CNA burden outperforms mutation-based metrics and can stratify risk even in tumors traditionally considered lower-risk. The results highlight the translational potential of CNA-derived biomarkers for precision risk assessment in underserved patient populations. Collectively, this thesis demonstrates that computational oncology tools can transform complex molecular data into clinically interpretable knowledge, improving diagnostic accuracy and enabling more personalized therapeutic strategies. By focusing on approaches compatible with formalin-fixed paraffin-embedded (FFPE) samples and existing clinical assays, the work ensures that methodological innovation is closely aligned with real-world implementation in molecular pathology. These contributions mark a step toward routine, multi-omic integration in precision cancer care, navigating the oncogenomic maze to support better outcomes for patients.