Algorithmic Tools for Mining High-Dimensional Cytometry Data

Abstract

The advent of mass cytometry has led to an unprecedented increase in the number of analytes measured in individual cells, thereby increasing the complexity and information content of cytometric data. Although this technology is ideally suited to the detailed examination of the immune system, the applicability of the different methods for analyzing such complex data is less clear. Conventional data analysis by manual gating of cells in biaxial dot plots is often subjective, time consuming, and neglectful of much of the information contained in a highly dimensional cytometric dataset. Algorithmic data mining has the promise to eliminate these concerns, and several such tools have been applied recently to mass cytometry data. We review computational data mining tools that have been used to analyze mass cytometry data, outline their differences, and comment on their strengths and limitations. This review will help immunologists to identify suitable algorithmic tools for their particular projects.

Figure 1. Principal component analysis of a 3-dimensional point cloud

The principal components are ordered by their capacity to explain the variance in the data.

Figure 2. Swiss roll

The artificial Swiss roll data provides an example of how linear dimensionality reduction can be misleading. Points A and B are proximate when considering their linear relationship in Euclidean space. However, when the entire structure of the data manifold is considered, the points are distant.

Figure 3. Hierarchical, agglomerative clustering

In agglomerative clustering, each point in multidimensional space is initially its own cluster. Then, the most phenotypically similar “clusters” merge into an agglomerative, parent cluster. This procedure then iterates until the target number of clusters is reached. Both SPADE and Citrus utilize agglomerative clustering.

Figure 4. Self-Organizing Map with Start Charts

A self-organizing map (SOM) is a type of artificial neural network used for clustering. In SOM-based clustering, each cell is assigned to a phenotypically similar node in the network. Characteristics of each node can be visualized via star charts. Star charts provide a way to visualize the mean intensities of the markers for all cells in the dataset assigned to a specific node. The height of each wedge within the star chart indicates the markers intensity (i.e. if the wedge reaches the border of the circle, the cells have a high expression of that marker).