Toward deterministic and semiautomated SPADE analysis

Abstract

SPADE stands for spanning-tree progression analysis for density-normalized events. It combines downsampling, clustering and a minimum-spanning tree to provide an intuitive visualization of high-dimensional single-cell data, which assists with the interpretation of the cellular heterogeneity underlying the data. SPADE has been widely used for analysis of high-content flow cytometry data and CyTOF data. The downsampling and clustering components of SPADE are both stochastic, which lead to stochasticity in the tree visualization it generates. Running SPADE twice on the same data may generate two different tree structures. Although they typically lead to the same biological interpretation of subpopulations present in the data, robustness of the algorithm can be improved. Another avenue of improvement is the interpretation of the SPADE tree, which involves visual inspection of multiple colored versions of the tree based on expression of measured markers. This is essentially manual gating on the SPADE tree and can benefit from automated algorithms. This article presents improvements of SPADE in both aspects above, leading to a deterministic SPADE algorithm and a software implementation for semiautomated interpretation. © 2017 International Society for Advancement of Cytometry.

Keywords: SPADE; deterministic.

Figure 1

Screenshots of the deterministic SPADE software: (a) main control window, (b) parameter setting window, and parameters used for the mouse bone marrow data.

Figure 1

Screenshots of the deterministic SPADE software: (a) main control window, (b) parameter setting window, and parameters used for the mouse bone marrow data.

Figure 2

Screenshot of SPADE visualization interface.

Figure 3

The first two automated partitioning suggestions. Boxplots show which markers support the suggested partitioning. Bubbles are drawn to visualize the partitioning. The tree is colored by the marker that contributes the most to the suggestion. In the upper-left tree, all nodes are colored. In the upper-right tree, only nodes involved in the suggested partitioning are colored.

Figure 3

The first two automated partitioning suggestions. Boxplots show which markers support the suggested partitioning. Bubbles are drawn to visualize the partitioning. The tree is colored by the marker that contributes the most to the suggestion. In the upper-left tree, all nodes are colored. In the upper-right tree, only nodes involved in the suggested partitioning are colored.

Figure 4

Semi-automated partitioning result: (a) visualization interface displaying the partitioning, (b) interactive boxplot and heatmap summarizing the marker expression combination of the annotations.

Figure 4

Semi-automated partitioning result: (a) visualization interface displaying the partitioning, (b) interactive boxplot and heatmap summarizing the marker expression combination of the annotations.

Figure 5

Robustness of deterministic SPADE and tree partitioning algorithms. 100 random subsamples of the mouse bone marrow dataset were analyzed, with 9 automatically generated annotations in each analysis. In the tSNE visualization, the centers of the 9*100 annotations formed 9 groups corresponding to distinct populations, which showed the robustness of the annotations.