GigaSOM.jl: High-performance clustering and visualization of huge cytometry datasets

Abstract

Background: The amount of data generated in large clinical and phenotyping studies that use single-cell cytometry is constantly growing. Recent technological advances allow the easy generation of data with hundreds of millions of single-cell data points with >40 parameters, originating from thousands of individual samples. The analysis of that amount of high-dimensional data becomes demanding in both hardware and software of high-performance computational resources. Current software tools often do not scale to the datasets of such size; users are thus forced to downsample the data to bearable sizes, in turn losing accuracy and ability to detect many underlying complex phenomena.

Results: We present GigaSOM.jl, a fast and scalable implementation of clustering and dimensionality reduction for flow and mass cytometry data. The implementation of GigaSOM.jl in the high-level and high-performance programming language Julia makes it accessible to the scientific community and allows for efficient handling and processing of datasets with billions of data points using distributed computing infrastructures. We describe the design of GigaSOM.jl, measure its performance and horizontal scaling capability, and showcase the functionality on a large dataset from a recent study.

Conclusions: GigaSOM.jl facilitates the use of commonly available high-performance computing resources to process the largest available datasets within minutes, while producing results of the same quality as the current state-of-art software. Measurements indicate that the performance scales to much larger datasets. The example use on the data from a massive mouse phenotyping effort confirms the applicability of GigaSOM.jl to huge-scale studies.

Keywords: Julia; clustering; dimensionality reduction; high-performance computing; self-organizing maps; single-cell cytometry.

Figure 1

Architecture of GigaSOM.jl. Top: Data distribution process divides the available FCS files into balanced slices; individual workers retrieve their respective slice data using a shared storage. Bottom:The SOM learning and visualization processes require only a minimal amount of data to be transferred between the master and worker nodes, consisting of the relatively small codebook in the case of SOM learning (blue arrows) and pre-rasterized graphics in the case of visualization (green arrows).

Figure 2

Comparison of GigaSOM.jl results with manual gating of the Levine32 dataset. The confusion matrix is normalized in rows, showing the ratio of cells in each aggregate of GigaSOM-originating clusters that matches the cell types from manual analysis. Darker color represents better match. The mean F1 score is comparable to FlowSOM. A more comprehensive comparison is available in Supplementary Fig. S1.

Figure 3

Performance dependency of distributed algorithms in GigaSOM on data dimensionality, SOM size, and number of available workers. Data processing performance is displayed as normalized to median speed in cells per second (c/s).

Figure 4

Effect of data-indexing structures on GigaSOM performance. The plotted points show relative speedup of the algorithms utilizing kd-trees (horizontal axis) and ball-trees (vertical axis) compared with brute-force neighbor search. Baseline (1× speedup) is highlighted by thick grid lines—a point plotted in the upper right quadrant represents a benchmark measurement that showed speedup for both kd-trees and ball-trees, upper left quadrant contains benchmark results where ball-trees provided speedup and kd-trees slowed the computation down, etc.

Figure 5

Raw IMPC Spleen T-cell dataset, processed by GigaSOM.jl and embedded by the Julia implementation of EmbedSOM. The figure shows an aggregate of 1,167,129,317 individual cells. Expression of 3 main markers is displayed in combination as mixed colors: CD8 in red, CD4 in green, and CD161 in blue. A more detailed, annotated version of the visualization is available in Supplementary Fig. S4.