D-AREdevil: a novel approach for discovering disease-associated rare cell populations in mass cytometry data

Résumé 0

Background: The advances in single-cell technologies such as mass cytometry provides increasing resolution of the complexity of cellular samples, allowing researchers to deeper investigate and understand the cellular heterogeneity and possibly detect and discover previously undetectable rare cell populations. The identification of rare cell populations is of paramount importance for understanding the onset, progression and pathogenesis of many diseases. However, their identification remains challenging due to the always increasing dimensionality and throughput of the data generated. Aim: This study aimed at implementing a straightforward approach that efficiently supports a data analyst to identify disease-associated rare cell populations in large and complex biological samples and within reasonable limits of time and computational infrastructure. Methods: We proposed a novel computational framework called D-AREdevil (disease- associated rare cells detection) for cytometry datasets. The main characteristic of our computational framework is the combination of an anomaly detection algorithm (i.e. LOF, or FiRE) that provides a continuous score for individual cells with one of the best performing and fastest unsupervised clustering methods (i.e. FlowSOM). In our approach, the LOF score serves to select a set of candidate cells belonging to one or more subgroups of similar rare cell populations. Then, we tested these subgroups of rare cells for association with a patient group, disease type, clinical outcome or other characteristic of interest. Results: We reported in this study the properties and implementation of D-AREdevil and presented an evaluation of its performances and applications on three different testing datasets based on mass cytometry data. We generated data mixed with one or more known rare cell populations at varying frequencies (below 1%) and tested the ability of our approach to identify those cells in order to bring them to the attention of the data analyst. This is a key step in the process of finding cell subgroups that are associated with a disease or outcome of interest, when their existence and identification is not previously known and has yet to be discovered. Conclusions: We proposed a novel computational framework with demostrated good sensitivity and precision in detecting target rare cell poopulations present at very low frequencies in the total datasets (