Page 301 - Contributed Paper Session (CPS) - Volume 4

CPS2233 Sharon Lee

A hierarchical mixed effects model for batch
cytometry data
Sharon X. Lee
School of Mathematical Sciences, University of Adelaide, South Australia, Australia

Abstract
Flow cytometry is an important tool in the diagnosis and monitoring of
immunological diseases such as lymphomas, leukaemia, and AIDS. It is
frequently used in immunological research, pre-clinical trials, and clinical
diagnosis. However, these data are challenging to model and analyze due to
the large number of observations and the inherent structure of the batch of
samples. Moreover, it is known that they typically exhibit non-normal features
such as asymmetry and heavy-tailedness. This paper considers the problem of
jointly modelling multiple cytometry data that comes from the same batch. In
particular, one of the aims is for the model to provide an automated
segmentation of the data. To achieve this, we adopt a hierarchical mixture
model approach to provide a probabilistic clustering of the data, together with
skew component distributions to cater for non-normal clusters. Furthermore,
our tool is designed to handle inter-data variations via the incorporation of a
random effects model. Examples from real cytometry experiments will be used
to demonstrate the effective od our approach.

Keywords
cytometry; mixed model; mixture model; clustering; skewness

1. Introduction
Flow cytometry is a powerful tool for characterizing single cell properties.
It is routinely used in both clinical and research immunology. Its ability to study
particles at the single-cell level renders it widely useful in many biomedical
fields. After staining with fluorophore-conjugated antibodies (or markers), the
sample is placed in a flow cytometer where cells are passed through a laser
beam one at a time. The light emerging from each cell are captured and
quantitated by different detectors. Modern cytometers can measure up to 30
markers simultaneously at a rate of 10,000 cells per second. This generates
datasets of massive size in a high-throughput manner.
A critical part of the analysis of flow cytometry data is the segmentation of
cells into different cell populations according to their properties. This task is
currently carried out manually where an analyst would visually discriminate
between different clusters or groups of points based on sequential bivariate
projections of the data. Not only is this process laborious and error-prone, but

290 | I S I W S C 2 0 1 9