GiniClust: detecting rare cell types from single-cell gene expression data with Gini index

Abstract

High-throughput single-cell technologies have great potential to discover new cell types; however, it remains challenging to detect rare cell types that are distinct from a large population. We present a novel computational method, called GiniClust, to overcome this challenge. Validation against a benchmark dataset indicates that GiniClust achieves high sensitivity and specificity. Application of GiniClust to public single-cell RNA-seq datasets uncovers previously unrecognized rare cell types, including Zscan4-expressing cells within mouse embryonic stem cells and hemoglobin-expressing cells in the mouse cortex and hippocampus. GiniClust also correctly detects a small number of normal cells that are mixed in a cancer cell population.

Keywords: Clustering; Gini index; RNA-seq; Rare cell type; Single-cell analysis; qPCR.

Fig. 1

Coparison between Gini index and Fano factor in detecting differentially expressed genes. a Scaled density plot of the expression levels of genes X (red) and Y (blue). The proportion of the minor cell type is 50 %. b The Lorenz curve for genes X (red) and Y (blue). The proportion of the minor cell type is 50 %. c, d Same as (a, b), except the proportion of the minor cell type is changed to 1e-5. e Fano factor for genes X and Y for varying proportions of the minor cell type (1/1 M stands for one in one million). f Gini index for genes X and Y for varying proportions of the minor cell type

Fig. 2

Overview of the GiniClust pipeline. Details are described in Methods

Fig. 3

GiniClust uncovers rare cell types from the qPCR dataset. a Relationship between the raw Gini index and the log2-transformed maximum expression level. Selected genes with high normalized Gini index values are labeled as red dots. b Overlap between the selected high Gini genes and differentially expressed genes. c t-SNE visualization of the data. Cells are color-coded based on the GiniClust cluster membership. d t-SNE visualization of the same data as in c. Cells are color-coded based on the actual lineage. e Expression levels of representative genes for MASC (n = 24), ISC (n = 23), and other cells (n = 1916). Gene expression levels are normalized as percentage of the corresponding maximum values. f Comparison between GiniClust and RaceID in detection of ISC and MASCs in the mixture of cells

Fig. 4

GiniClust identifies a Zscan4-enriched rare cluster from mouse embryonic stem cells. a Relationship between the raw Gini index and the log2-transformed maximum expression level. Selected genes with high normalized Gini index values are labeled as red dots. b t-SNE visualization of the data. Cells are color-coded based on the GiniClust cluster membership. Inset shows a zoomed-in region around the rare cell cluster. c Overlap between the selected high Gini genes and upregulated genes in Cluster 2. d Expression pattern of representative genes (Tcstv1, Dcdc2c, Zscan4f, Zscan4d) in Cluster 2 (n = 3, left panels) compared to Cluster 1 (n = 2505, right panels). Each bar represents a single cell

Fig. 5

GiniClust identifies a rare cluster in glioblastoma samples. a Relationship between the raw Gini index and the log2-transformed maximum expression level. Selected genes with high normalized Gini index values are labeled as red dots. b t-SNE visualization of the data. Cells are color-coded based on the GiniClust cluster membership. c Overlap between the selected high Gini genes and upregulated genes in Cluster 2. d Expression pattern of representative genes (CLDN11, MBP, PLP1, KLK6) in Cluster 3 (n = 9, left panels) compared to Cluster 1 (n = 261, right panels). Each bar represents a single cell

Fig. 6

GiniClust identifies a rare cell type in mouse cortex and hippocampus. a Relationship between the raw Gini index and the log2-transformed maximum expression level. Selected genes with high normalized Gini index values are labeled as red dots. b t-SNE visualization of the data. Cells are color-coded based on the GiniClust cluster membership. c Overlap between the selected high Gini genes and upregulated genes in Cluster 4. d Expression pattern of representative genes (Hba-a2, Hbb-b2, Hbb-bs) in Cluster 4 (n = 3, left panels) compared to Cluster 1 (n = 1842, right panels). Each bar represents a single cell. The expression levels of Hba-a2 shown here represent the sum of the levels of Hba-a2_loc1 and Hba-a2_loc2 in the original paper