From signal to survival

Résumé fourni par la source

Acute myeloid leukemia (AML) is a heterogeneous myeloid neoplasm characterized by uncontrolled proliferation of immature myeloid blasts. Although most patients treated with intensive chemotherapy can achieve a state of morphological remission, residual leukemic cells may remain in the bone marrow and are associated with relapse. Different assays have been developed to detect this measurable residual disease (MRD). One of these techniques, flow cytometry, is widely applicable for AML patients. However, data analysis is based on manual gating, a laborious and potentially subjective process that hampers widespread adoption of flow cytometry-based assays outside specialized centers. This thesis is split into three parts (Part I - III), each focusing on crucial aspects of flow cytometry data analysis in the translational setting. In Part I, we focus on aspects related to pre-processing and quality control of flow cytometry data. The broad implementation of flow cytometry in translational research now results in datasets that can comprise thousands of measurements. For such datasets, manual quality control is close to impossible. To address this, we developed a software package called CytoScan in Chapter 2 for visualization and inspection of large cytometry cohorts, with a particular emphasis on anomaly detection. Limitations in flow cytometry equipment also often require the distribution of different markers across multiple tubes. When shared backbone markers are present, these can be used to calculate or impute missing markers and allow for a combined analysis of complete set of markers for all cells. In Chapter 3, we show that the data generated by such imputation methods is of insufficient quality for use in clinical settings. In Part II, we focus on computational measurable residual disease (cMRD) assessment. In Chapter 4, we review the current landscape of computational methods for AML-MRD assessment. We highlight how the heterogeneity of immunophenotypic patterns and the rarity of leukemic blasts in the MRD setting pose challenges for existing algorithms. Next, in Chapter 5, we propose a novel methodology based on Gaussian mixture models (GMMs) that allows for a fully automated MRD assessment in three seconds per sample. By combining an element of supervised learning for blast identification with GMM-based novelty detection, we effectively reproduce the philosophy behind the existing diagnostic workflow. In Chapter 6, we clinically validate this method in a retrospective cohort of 399 AML patients and find that like manual gating, cMRD is prognostic for outcomes of AML patients. Computational cytometry data analysis is not restricted to automating established manual gating procedures but also benefits from unsupervised analyses that capture distinct characteristics in the data. In Part III, we extend the default GMM architecture to allow for improved modeling of heterogeneous patient cohorts that are typical for AML (Chapter 7). We apply this new algorithm, MSGMM, in an extensive study of the immunophenotypic characteristics of newly diagnosed AML patients in Chapter 8. Taken together, this thesis shows how computational methods can aid, automate and augment the analysis of immunophenotypic characteristics in AML, enabling the translation from measurements (signal) to prognosis and outcomes (survival).

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.