Towards good representations of single-cell protein expression data

Abstract

Many diseases including infections and cancer can evoke an immune response that is detectable as changes in the immune cell composition in blood. In cancers originating in the immune system, the immune cell composition can also change due to uncontrolled growth of the malignant cells, shifting the normal balance between immune cell types. One example is acute myeloid leukemia (AML), which affects a precursor of several immune cell types found in blood including monocytes and neutrophils. Single-cell protein expression measurements obtained with CyTOF offer a powerful means of studying immune cell composition in blood, and this thesis concerns the analysis of such data. We specifically consider the problem of converting such data - which describe which proteins are expressed on each analyzed cell - to a representation that reveals the immune cell composition. We also study how to obtain representations that are well suited as inputs to algorithms for prediction of treatment outcome and survival. The focus will be on unsupervised clustering, and we propose a novel semi-supervised clustering algorithm and compare its performance with other methods.