Rapid cell population identification in flow cytometry data

Abstract

We have developed flowMeans, a time-efficient and accurate method for automated identification of cell populations in flow cytometry (FCM) data based on K-means clustering. Unlike traditional K-means, flowMeans can identify concave cell populations by modelling a single population with multiple clusters. flowMeans uses a change point detection algorithm to determine the number of sub-populations, enabling the method to be used in high throughput FCM data analysis pipelines. Our approach compares favorably to manual analysis by human experts and current state-of-the-art automated gating algorithms. flowMeans is freely available as an open source R package through Bioconductor.

Figure 1

An example of finding the change point using segmented regression. The chosen solution (shown in red) consists of 6 populations.

Figure 2

The number of clusters selected by manual analysis and the three algorithms for the (a)GvHD and (b)DLBCL datasets.

Figure 3

Agreement between F-measures of flowMeans and either flowMerge(a,b) or FLAME(c,d) on GvHD(a,c) and DLBCL(b,d) datasets. The cell populations for the samples indicated with red X’s in panels (a)-(d) are shown in respective panels in Figure 4. The dashed line is the agreement line (i.e., y = x) that indicates where the performance of the two algorithms is equal. The correlation coefficient (CC) and concordance correlation coefficient (CCC) are shown as legends.

Figure 4

Panels (a)-(d) illustrate the cell populations found by flowMeans, flowMerge, and FLAME for the samples shown with red X’s in respective panels in Figure 3. In this figure, the >90th percentiles of each cluster are visualized to make the boundaries more robust after projection to a two dimensional scatter plot. Therefore the populations might be different from the real distributions on the margins. The pink cluster in panel (d) is a multi-modal population with 2 high-density regions. In every panel, colors of each solution are matched with the solution with the maximum number of clusters.