Outils bio-informatiques pour l'analyse de données unicellulaires : identification de sous-populations résistantes et appui aux décisions thérapeutiques dans le cancer de l'ovaire

Although many patients with high-grade serous ovarian cancer (HGSOC) initially respond to platinum-based chemotherapy, eventually most patients experience disease recurrence and develop resistance to treatment, resulting in poor clinical outcomes. Single-cell technologies offer a powerful means to dissect intratumoral heterogeneity and identify distinct subpopulations of cancer cells that may drive resistance. However, the complexity and high dimensionality of this data requires sophisticated computational approaches for robust analysis. In this thesis, we pursued two main objectives with a key eye on identifying and developing computational approaches for biologically meaningful analysis. First, we aimed to identify clinically resistant subpopulations of cancer cells. Second, we sought to select drugs that selectively target these resistance-associated populations. This thesis begins in the first chapter with a comparative study of Cancer Antigen (CA)-125-based chemosensitivity scores as an alternative metric to the platinum-free interval, a commonly used clinical measure of resistance in HGSOC. Unlike platinum-free interval (PFI), which can be confounded by factors such as follow-up variability, CA-125-based scores offer a potentially more direct assessment of chemosensitivity in patients. Our analysis showed that most scores were significantly associated with PFI in patients treated with neoadjuvant chemotherapy, but exhibited weaker associations with other relevant clinical variables. In contrast, the associations were generally poorer in patients undergoing primary debulking surgery before adjuvant chemotherapy, underscoring the need for more robust metrics across treatment settings. In the second chapter, we present a comprehensive bioinformatics pipeline designed to detect chemotherapy-resistant cell subpopulations from single-cell RNA-sequencing (scRNA-seq) and clinical data and to guide treatment selection. To address our first objective, we used non-negative matrix factorization (NMF) to model the expression profile of each cell as a weighted sum over elementary gene signatures. We then assessed each signature's association with resistance using clinical variables and differences in cell proportions expressing the signature (pre- to post-chemotherapy). Additionally, external validation is performed in bulk RNA-seq data. For our second objective, we use data from large-scale drug screenings to identify existing drugs that target each resistance-associated signature. We applied our framework to a unique scRNA-seq dataset of 57 samples collected at different time points during treatment from 42 HGSOC patients. The resulting signatures and predicted drugs were experimentally validated in patient derived organoids and xenograft models. While NMF has proven valuable in identifying principal gene expression programs, it remains challenging to identify subtler gene expression patterns, particularly in complex clinical datasets from multiple cancer patients, which are characterised by extensive inter- and intra-patient heterogeneity. In this context, patient or sample level annotations are often available but ignored at this stage of data analysis. In cases where one searches for expression programs that are ultimately associated with specific phenotypes, like resistance, this association can be used to constrain NMF, and therefore facilitate the identification of such phenotype-related expression programs. In the last chapter, we propose a supervised NMF approach that implements this idea by extending the standard NMF algorithm by a classifier module that enables the identification of general expression programs as well as programs that are associated with a phenotype. In summary, this research developed analytical pipelines and tools towards establishing personalized treatment strategies for platinum-resistant HGSOC tumors. While focused on ovarian cancer, the approaches are broadly applicable to other cancer types.