RESOLVING TUMOR HETEROGENEITY THROUGH HIGH RESOLUTION MULTI-MODAL TRANSCRIPTOMICS: INTEGRATIVE SINGLE-CELL ATLASES AND GENERATIVE SPATIAL MODELING

Understanding tumor heterogeneity is essential for improving diagnosis, treatment, and drug development in oncology. This thesis explores how cutting-edge transcriptomic technologies and new computational frameworks can be used to dissect, interpret, and model the cellular complexity of solid tumors, with a particular focus on two case studies namely high-grade serous ovarian cancer (HGSC) and glioblastoma multiforme (GBM). The first part of the thesis focuses on building and leveraging a harmonized single-cell transcriptomic atlas (OvCA) constructed from patient-derived samples across multiple tissue sites and treatment stages. Following rigorous feature selection, integration, and cell state annotation pipelines, the atlas reveals conserved and tissue-specific programs across malignant and stromal compartments, uncovers treatment-associated shifts, and identifies robust tumor–microenvironment interactions. These insights are then extended to new samples thanks to the deep learning model derived from data integration allowing for the iterative expansion of the atlas and applications in in vitro experimental settings. Next, I apply this framework to assess the biological fidelity of in vitro models. Using single-cell RNA-seq, I profiled 2D and 3D cultures grown in the presence or absence of ascitic fluid (AF). AF supplementation enhanced proliferation, preserved minor tumor populations, and reshaped the composition of cancer-associated fibroblasts (CAFs). The transcriptomic landscape of these models was further integrated into the OvCA atlas to validate their biological relevance and investigate tumor cell plasticity under different culture conditions. The second arm of the thesis addresses the spatial dimension of tumor heterogeneity. While current spatial transcriptomics platforms are limited in gene coverage, they offer unprecedented insight into the spatial architecture of tissues. To exploit this, I contributed to the development and application of MintFlow, a computational framework designed to uncover spatially relevant transcriptional programs by disentangling cell-intrinsic expression from microenvironment-induced signals. Applying MintFlow to spatial data from GBM, I identified spatially localized malignant programs, including hypoxia-dependent angiogenic and infiltrative states. These results advance our ability to interrogate tumor ecosystems in situ, offering mechanistic insight into spatial patterning and its role in disease progression. In sum, this work highlights the value of integrative multi-modal transcriptomics to dissect tumor heterogeneity in both molecular and spatial dimensions. By combining high-resolution atlases, in vitro modelling, and generative spatial inference, this thesis provides both a conceptual and practical foundation for future applications in systems oncology and AI-driven translational research.