Evaluation of methods to assign cell type labels to cell clusters from single-cell RNA-sequencing data

Identification of cell type subpopulations from complex cell mixtures using single-cell RNA-sequencing (scRNA-seq) data includes automated steps from normalization to cell clustering. However, assigning cell type labels to cell clusters is often conducted manually, resulting in limited documentation, low reproducibility and uncontrolled vocabularies. This is partially due to the scarcity of reference cell type signatures and because some methods support limited cell type signatures.

University of Toronto researchers benchmarked five methods representing first-generation enrichment analysis (ORA), second-generation approaches (GSEA and GSVA), machine learning tools (CIBERSORT) and network-based neighbor voting (METANEIGHBOR), for the task of assigning cell type labels to cell clusters from scRNA-seq data. They used five scRNA-seq datasets: human liver, 11 Tabula Muris mouse tissues, two human peripheral blood mononuclear cell datasets, and mouse retinal neurons, for which reference cell type signatures were available. The datasets span Drop-seq, 10X Chromium and Seq-Well technologies and range in size from ~3,700 to ~68,000 cells.

Thier results show that, in general, all five methods perform well in the task as evaluated by receiver operating characteristic curve analysis (average area under the curve (AUC) = 0.91, sd = 0.06), whereas precision-recall analyses show a wide variation depending on the method and dataset (average AUC = 0.53, sd = 0.24). The researchers observed an influence of the number of genes in cell type signatures on performance, with smaller signatures leading more frequently to incorrect results.

Schematic of a process to benchmark automated cell type prediction methods

Two inputs are needed by automated cell type prediction methods ( A– C). ( A) a matrix with the average expression of each gene x for each cell cluster y ( Ě _xy). ( B, C) cell type gene marker signatures can be provided as either gene sets (lists of gene identifiers, B) or numeric gene expression profiles ( C). ( D) Gene sets can be manually compiled from literature and are used for methods like GSEA, GSVA or ORA, whereas gene-expression profiles are measurements from microarrays, bulk- or single-cell RNA-sequencing (scRNA-seq) experiments and are used by methods like CIBERSORT and METANEIGHBOR. ( E) Automated cell type prediction methods produce a matrix of cell type prediction scores for each cell cluster. ( F) Some authors of scRNA-seq studies assign cell type labels manually to cell clusters using local expertise or orthogonal experiments such as fluorescence activated cell sorting. These annotations can be used as a gold standard to benchmark automated cell type predictions. ( G) Cell type prediction scores (from E) for cell clusters are concatenated into a single vector and known cell cluster annotations (from F) are added. The resulting matrix is used to assess the performance of cell type prediction methods by receiver operating characteristic (ROC) curve and precision-recall (PR) curve analyses varying over the prediction scores for all cell clusters in a dataset ( H). ( I) Robustness of cell type prediction methods can be analysed by gradually subsampling gene markers from cell type gene expression signatures ( B or C) and repeating procedures of ( D– H) to obtain distributions of the area under the curve (AUC) for ROC (ROC AUC) and PR (PR AUC) curves, which are shown as violin plots. We hypothesized that some prediction methods are more robust than others to the proportion of gene markers subsampled from cell type gene expression signatures.

GSVA was the overall top performer and was more robust in cell type signature subsampling simulations, although different methods performed well using different datasets. METANEIGHBOR and GSVA were the fastest methods. CIBERSORT and METANEIGHBOR were more influenced than the other methods by analyses including only expected cell types.

Availability – The researchers provide an extensible framework that can be used to evaluate other methods and datasets at https://github.com/jdime/scRNAseq_cell_cluster_labeling.