Identifying predictors of survival in patients with leukemia using single-cell mass cytometry and machine learning

Abstract

Mass cytometry by time-of-flight (CyTOF) is an emerging technology allowing for in-depth characterisation of cellular heterogeneity in cancer and other diseases. However, computational identification of cell populations from CyTOF, and utilisation of single cell data for biomarker discoveries faces several technical limitations, and although some computational approaches are available, high-dimensional analyses of single cell data remains quite demanding. Here, we deploy a bioinformatics framework that tackles two fundamental problems in CyTOF analyses namely: a) automated annotation of cell populations guided by a reference dataset, and b) systematic utilisation of single cell data for more effective patient stratification. By applying this framework on several publicly available datasets, we demonstrate that the Scaffold approach achieves good tradeoff between sensitivity and specificity for automated cell type annotation. Additionally, a case study focusing on a cohort of 43 leukemia patients, reported salient interactions between signalling proteins that are sufficient to predict short-term survival at time of diagnosis using the XGBoost algorithm. Our work introduces an automated and versatile analysis framework for CyTOF data with many applications in future precision medicine projects. Datasets and codes are publicly available at: https://github.com/dkleftogi/singleCellClassification