Abstract Body

Flow and mass (CyTOF) cytometry are high-throughput technologies quantifying multiple surface and intracellular markers at the level of a single cell. Improvements of these technologies allow to describe millions of individual cells from a single blood sample according to several dozens of markers (up to 2^30 cell populations with 30 markers). This generate high-dimensional datasets, whose manual analysis, called manual gating, is highly time-consuming and poorly reproducible. We have developed 2 machine learning approaches to perform automatic gating without human intervention.

The first method is a new Bayesian nonparametric approach (NPflow) with Dirichlet process mixture (DPM) of multivariate skew t-distributions to perform model-based clustering of flow-cytometry data. DPM models directly estimate the number of cell populations from the data, avoiding model selection issues, and skew t-distributions provides robustness to outliers and non-elliptical shape of cell populations. To accommodate repeated flow-cytometry measurements, such as in a clinical trial, a sequential strategy relying on a parametric approximation of the posterior is also proposed (NPflow seq). The second one (cytometree) is based on the construction of a binary tree, whose nodes represents cellular sub-populations. At each node, a binary split between different cellular populations is done according to the normalized difference of Akaike Information Criteria (AIC) between the two normal mixture models considering either one or two possible sub-populations. Post-processing of the tree structure and derived populations allows us to automatically provide a complete annotation of the derived populations.

The good performance of the methods are shown on simulated data and on an experimental benchmark datasets (FLOWCAP1) as shown in Table 1, as well as in a real dataset from an HIV vaccine trial. Compared to other available approaches the new methods, especially cytometree, performed at the top position with the shortest runtime. Also, the F-measures>0.90 demonstrate the validity of the new methods in comparison with the gold standard (consensus of 8 experimentalists).

The constant increase of the number of markers available to characterize cell populations leads to an untractable situation with manual gating. However, improvements of machine learning approaches allow for the automatic analysis of cytometry samples, yielding the relative count for all possible cell populations.