FR-Match – robust matching of cell type clusters from single cell RNA sequencing data using the Friedman-Rafsky non-parametric test

Single cell/nucleus RNA sequencing (scRNAseq) is emerging as an essential tool to unravel the phenotypic heterogeneity of cells in complex biological systems. While computational methods for scRNAseq cell type clustering have advanced, the ability to integrate datasets to identify common and novel cell types across experiments remains a challenge.

Researchers from the J. Craig Venter Institute have developed a cluster-to-cluster cell type matching method—FR-Match—that utilizes supervised feature selection for dimensionality reduction and incorporates shared information among cells to determine whether two cell type clusters share the same underlying multivariate gene expression distribution. FR-Match is benchmarked with existing cell-to-cell and cell-to-cluster cell type matching methods using both simulated and real scRNAseq data. FR-Match proved to be a stringent method that produced fewer erroneous matches of distinct cell subtypes and had the unique ability to identify novel cell phenotypes in new datasets. In silico validation demonstrated that the proposed workflow is the only self-contained algorithm that was robust to increasing numbers of true negatives (i.e. non-represented cell types). FR-Match was applied to two human brain scRNAseq datasets sampled from cortical layer 1 and full thickness middle temporal gyrus. When mapping cell types identified in specimens isolated from these overlapping human brain regions, FR-Match precisely recapitulated the laminar characteristics of matched cell type clusters, reflecting their distinct neuroanatomical distributions.

FR-Match schematic and marker gene ‘barcodes’

(a) FR-Match cluster-to-cluster matching schematic diagram. Input data: query/new and reference datasets, each with cell-by-gene expression matrix and cell cluster membership labels. Step I: dimensionality reduction by selecting expression data of reference cell type marker genes from the query dataset. Here, we use the NS-Forest marker genes selected for the reference cell types. Step II: Cluster-to-cluster matching through the FR test. From left to right: multivariate data points of cell transcriptional profiles (colored by cell cluster labels) in a reduced dimensional (reference marker gene expression) space; construct a complete graph by connecting each pair of vertices (i.e. cells); find the minimum spanning tree that connects all vertices with minimal summed edge lengths; remove the edges that connect vertices from different clusters; count the number of disjoint subgraphs, termed ‘multivariate runs’ and denoted as RR⁠; calculate the FR statistic WW⁠, which has asymptotically a standard normal distribution. (b) ‘Barcodes’ of the cortical Layer 1 NS-Forest marker genes in four Layer 1 clusters. Heatmaps show marker gene expression levels of 30 randomly selected cells in each cell cluster. The ‘Marker’ column indicates if the gene is a marker gene of the cluster or not (1 = yes, 0 = no).

Availability – An R package and Shiny application are provided at https://github.com/JCVenterInstitute/FRmatch for users to interactively explore and match scRNAseq cell type clusters with complementary visualization tools.